A hybrid model for day-ahead electricity price forecasting:

Combining fundamental and stochastic modelling

Mira Watermeyera,∗ , Thomas Möbiusb,∗∗ , Oliver Grothea and Felix Müsgensb

a Karlsruhe Institute of Technology (KIT), Institute for Operations Research, Karlsruhe, 76131, Germany
arXiv:2304.09336v1 [econ.EM] 18 Apr 2023

b Brandenburg University of Technology (B-TU), Chair of Energy Economics, Cottbus, 03046, Germany

ARTICLE INFO
ABSTRACT

Keywords:
The accurate prediction of short-term electricity prices is vital for effective trading strategies,
Electricity price forecasting power plant scheduling, profit maximisation and efficient system operation. However, uncer-
tainties in supply and demand make such predictions challenging. We propose a hybrid model
Hybrid model that combines a techno-economic energy system model with stochastic models to address this
Energy system modelling challenge. The techno-economic model in our hybrid approach provides a deep understanding
of the market. It captures the underlying factors and their impacts on electricity prices, which
Stochastic modelling is impossible with statistical models alone. The statistical models incorporate non-techno-
economic aspects, such as the expectations and speculative behaviour of market participants,
Error improvement through the interpretation of prices. The hybrid model generates both conventional point
Probabilistic forecasting predictions and probabilistic forecasts, providing a comprehensive understanding of the market
landscape. Probabilistic forecasts are particularly valuable because they account for market
uncertainty, facilitating informed decision-making and risk management. Our model delivers
state-of-the-art results, helping market participants to make informed decisions and operate their
systems more efficiently.

1. Introduction
Accurate forecasting in the energy sector is crucial for multiple stakeholders, including industry practitioners,
researchers and policymakers. The effectiveness of financial and operational decisions and regulatory interventions
depends on accurate predictions of future developments in relevant areas. As a result, the forecasting of electricity
prices has become a key area of focus [Weron, 2014]. With companies facing increasingly intense competition due to
deregulation and liberalisation in the electricity sector, day-ahead price forecasts and insight into the next day’s market
situation are essential to the development of bidding strategies and production plans that maximise a company’s profit
margins and ensure a reliable grid operation. Quantifying uncertainty has become increasingly important in recent
years due to the growth of renewable energies and the need to integrate them alongside an increase in infrastructural
challenges and fluctuating commodity prices, raising uncertainty in the energy market [Nowotarski and Weron, 2018,
Hong et al., 2016, 2020]. Probabilistic forecasts help in the planning and operation of energy systems, allowing for the
assessment of uncertainty and the development of future strategies against the background of various probable future
events [Amjady and Hemmati, 2006].
Our paper presents a novel, open-source hybrid model that forecasts day-ahead electricity prices punctually and
probabilistically by combining two main methodological streams: techno-economic energy system modelling and
stochastic modelling. Techno-economic models are fundamental energy system models that determine (partial) market
equilibria through the bottom-up optimisation of an energy system. They can explain actual developments and reflect
structural breaks by identifying techno-economic interdependencies in energy markets. However, when estimating
[email protected] (.M. Watermeyer); [email protected] (T. Möbius)
https://www.as.ior.kit.edu/ (.M. Watermeyer); https://www.b-tu.de/fg-energiewirtschaft/ (T. Möbius)
ORCID (s):

Watermeyer et al. Page 1 of 38

 A hybrid model for day ahead electricity price forecasting

prices in the short term (e.g., day-ahead, intraday), these models exhibit larger and more systematic errors than other
model classes. Stochastic models, on the other hand, learn from history and are developed and trained with historical
data, enabling them to capture fluctuations and uncertainties in the market, especially in the short term. They offer high
flexibility and the ability to specify forecast ranges and distributions. Still, they can only capture structural breaks and
changes in external influences ex-post due to their dependence on historical data.
Our proposed hybrid model combines the strengths of techno-economic energy system models and stochastic models
to develop a more robust and accurate approach to forecasting electricity prices on the day-ahead market. The model
retains the structural statements of techno-economic energy system models – and, thus, insights into the driving market
mechanisms – while incorporating stochastic short-term structures and distribution functions to account for uncertainty.
The model uses state-of-the-art methods to generate point and probabilistic price forecasts. These probabilistic price
forecasts, with probabilities for each potential price scenario, are increasingly valuable to the industry (e.g., when
assessing the probabilities of negative prices or when assessing the overall risk level of the price forecast).
The model is schematically illustrated in Figure 1. It employs a rolling-window approach. In each iteration, it forecasts
day-ahead prices exclusively through the use of data known prior to the day-ahead market’s closure, accurately
reflecting the knowledge of stakeholders making decisions in these markets. The model is repeatedly applied each
day (𝑑𝑛 ) to generate forecasts for the following 24 hours of the day-ahead market. Each daily forecast includes four
steps – stochastic data pre-processing, parameter density forecast, energy system optimisation and stochastic data post-
processing – to produce point and probabilistic forecasts.

Figure 1: Scheme of the hybrid model.

The first step, stochastic data pre-processing, aims to improve the accuracy of input data in advance of the energy
system optimisation step and generates the basis for the parameter density forecast step. In the second step, the
parameter density forecast generates prediction intervals for selected input parameters of the third step, energy system
optimisation, enabling us to account for uncertainty in the operational decisions of market participants. This step
considers the improved input data from the first and second steps. It forms a stochastic optimisation model that
minimises total system costs, identifies the equilibrium between supply and demand and determines the hourly marginal
system costs, which can be interpreted as price estimators.
These price estimators are the initial values in the fourth and final step, stochastic data post-processing. The errors of
the price estimators are mapped using a multidimensional model in which seasonal effects and structures are captured
using a combination of univariate and multivariate approaches, resulting in an enhanced price forecast. By modelling
and improving the price forecast error, the model calculates forecast intervals and probability densities for the forecast
prices, providing a quantification of uncertainty. Finally, our model can combine the strengths of both method classes
to achieve excellent state-of-the-art price forecasts, including both point and probabilistic forecasts, to capture the
stochastic uncertainty of the market.
This paper contributes to the literature in three main ways. First, it presents a novel hybrid model that provides a
general framework to combine techno-economic and stochastic energy models. Since the model’s source code and
algebra are available in full online, other researchers can apply our methodology and extend it to different time periods
and electricity markets around the world. Second, it proves that techno-economic energy system models can contribute

Watermeyer et al. Page 2 of 38

 A hybrid model for day ahead electricity price forecasting

to short-term price forecasting, especially when paired with stochastic models for the sake of error improvement.
Our hybrid model delivers highly accurate day-ahead price forecasts on top of the insights that techno-economic
models provide. We demonstrate its value with an empirical analysis based on European data with a focus on Germany,
the largest European electricity market. Third, our hybrid model provides probabilistic forecasts in addition to point
forecasts, enabling power plant operators to, for example, quantify the likelihood of prices becoming negative at any
given hour.
The remainder of this chapter is organised as follows. Section 2 reviews the existing literature. Afterwards, we provide
all of the information necessary to use our model and replicate this study in Section 3. Section 4 describes our
methodology. Following, Section 5 presents and evaluates the results of our hybrid model, while Section 6 closely
analyses the individual model steps. Finally, Section 7 offers some concluding remarks.

2. Literature
Research on electricity prices has garnered the interest of many scholars due to the complexities and extraordinary
challenges of achieving high accuracy in forecasting. They have developed and refined numerous methodological
approaches to achieve high accuracy and adapt to changes in the electricity market. The number of relevant publications
has increased rapidly over the last two decades.
Weron [2014] offers a detailed review of several approaches to forecasting electricity prices, including the following
five model classes: multi-agent models, fundamental models, reduced-form models, statistical models, and artificial
intelligence models. Previously, Aggarwal et al. [2009] provide an overview of the methods used in electricity price
forecasting. However, their focus was on stochastic time-series, causal and artificial intelligence-based models. Weron
and Ziel [2019] and Hong et al. [2020] present a general review of and outlook on energy forecasting. The most recent
overview of forecasting theory and practice comes from Petropoulos et al. [2022], who provide an overview of a wide
range of theoretical models, methods, principles and approaches to preparing, organising and evaluating forecasts. In
addition, they provide several real-world examples of how these theoretical concepts are applied.
Many publications conduct time-series analysis based on time-series models, which are particularly suitable for short-
term electricity price forecasting. Time series models constitute a special subtype of regression model in which
target variables 𝑦 are represented, among other things, by past values of the time series as regressors 𝑥. They
include autoregressive moving-average (ARMA), generalised autoregressive conditional heteroscedastic (GARCH)
and Markov regime-switching (MS) models. Steinert and Ziel [2019], for example, develop an auto-regressive model
with 24 individual models – one for each hour of the day – that also incorporates electricity futures prices to produce
hourly electricity price forecasts. Nowotarski and Weron [2016] refer to the decomposition of values into different
components, which is common in time-series analysis, and show that the quality of time-series models benefits
greatly from decomposing a set of electricity prices into a long-term seasonal component and a stochastic component,
modelling them independently and combining their forecasts. In an extensive study, Ziel and Weron [2018] compare
two options for the type of time-series modelling used with high-frequency data sets. They compare models with
univariate model frameworks, with one of these models being constructed for the entire time series, featuring models
with multivariate model frameworks, in which each hour of a day is modelled separately and independently. This is
initiated by the organisation of electricity markets as day-ahead auctions, as in the U.S. or Europe. Their study shows
no clear dominance by one framework, suggesting that the combination of both modelling approaches could improve
predictive accuracy.
While Steinert and Ziel [2019], Nowotarski and Weron [2016] and Ziel and Weron [2018] focus on day-ahead electricity
prices in general, Christensen et al. [2012] use a nonlinear variant of the autoregressive conditional risk model to
predict price peaks, treating them collectively as a discrete-time point process, Eichler et al. [2014] an approach
based on the autoregressive conditional hazard model, and Manner et al. [2016] the mapping of inter-regional linkages
between different electricity markets in a dynamic multivariate binary choice model to predict electricity price spikes.
Garcia et al. [2005] develop a GARCH model to predict day-ahead electricity prices, while Hickey et al. [2012]
evaluate the accuracy of ARMAX-GARCH models in forecasting short-term prices in the U.S., determining that
model choice depends largely on location, horizon and regulation, with asymmetric power auto-regressive conditional
heteroskedastic (APARCH) models being more appropriate in deregulated markets and GARCH models being better for

Watermeyer et al. Page 3 of 38

 A hybrid model for day ahead electricity price forecasting

regulated markets. Bordignon et al. [2013] develop a linear regression model to account for relationships between prices
and various price drivers, using a time-varying parameter (TVR) and an MS model to capture peaks and discontinuities.
Other examples of applying MS models include Kosater and Mosler [2006] for the German market and Bierbrauer
et al. [2004] for the Nordic market. Notably, in a recent paper, Mari and Mari [2022] uses deep learning-based regime-
switching models to predict electricity prices.
Parameter-rich ARX models represent a special type of time-series model. The Lasso estimated autoregressive (LEAR)
model introduced by Uniejewski et al. [2016] is further developed by Lago et al. [2021].
To provide a set of best practices for evaluating future model developments in electricity price forecasting and
comparing state-of-the-art statistical and deep-learning methods, Lago et al. [2021] define a deep neuronal network
(DNN) and a LEAR model based on the latest findings. Together with various evaluation metrics, these models are
accessible in a Python toolbox to evaluate new algorithms. Accordingly, we compare our hybrid model with the
statistical benchmark model.
Deep-learning models (e.g., artificial neural network (ANN), DNN, long short-term memory (LSTM) network,
recurrent neural network (RNN), feed-forward neural network) are used in an increasing share of electricity price
forecasts. In addition to the cited benchmark model, Panapakidis and Dagoumas [2016], as an example, study
ANNs using different inputs and ANN typologies. These authors characterise such models as having comprehensive
functionality and a high degree of flexibility. In their analysis of the impact of different markets on one another,
Lago et al. [2018] develop a DNN that considers interconnected markets’ characteristics to boost forecasting accuracy.
Notably, they show that predicting the price of two markets simultaneously enhances forecast accuracy. Amjady [2006]
develops a fuzzy neuronal network that forecasts hourly electricity prices for the Spanish day-ahead market. Notably,
the combination of deep learning and time-series models can be found in the nonlinear autoregressive neural network
of Marcjasz et al. [2019].
To compare time-series and neural network models using external regressors, Lehna et al. [2022] use four different
forecasting approaches to the German day-ahead electricity market: a seasonal integrated autoregressive moving
average ((S)ARIMA(X)) model, an LSTM neural network, a convolutional neural network LSTM (CNN-LSTM) and
an extended two-stage multivariate vector autoregressive (VAR) model. While the LSTM model achieves the highest
average accuracy, the two-stage VAR model has advantages at shorter prediction horizons. A combination of both
methods outperforms each of the individual models in terms of accuracy.
The methods presented so far are fundamentally based on historical day-ahead electricity price time series. Another
approach entails using models that simulate the actions of individual market participants (agents). Qussous et al. [2022],
for example, developed an agent-based model to derive day-ahead prices and simulate the bidding strategies of market
participants. To evaluate their model, they reproduce day-ahead electricity prices in the 2016–2019 German bidding
zone. Compared to other techno-economic approaches to short-term electricity price forecasting, this agent-based
model achieves the highest accuracy. Consequently, we compare the hybrid model presented in this paper with this
model.
Due to their explanatory character in identifying an efficient market outcome and their comprehensive modelling of
the entire electricity system, techno-economic energy system models have been employed for the ex-post analysis of
electricity prices. Muesgens [2006a] and Borenstein et al. [2002] replicate day-ahead prices to assess the existence of
market power in Germany and the U.S., respectively. Keles et al. [2013] investigate on the importance of adequate wind
power feed-in time series to obtain better results in electricity price simulation. Hirth [2013] determine the market value
of renewables and Sensfuß et al. [2008] quantify the merit order effect computing day-ahead prices. The merit order
effect describes the displacement of fossil fuel generation by renewable energy sources due to their lower marginal costs
and the subsequent decline in total electricity costs. Pape et al. [2016] analyse to what extent day-ahead and intraday
electricity prices can be explained and represented by techno-economic energy system models. Notably, however, they
show that this method has significant weaknesses in explaining short-term electricity prices compared to other methods.
In contrast to ex-post analyses of electricity prices, which examine actual prices that have already occurred, this
article focuses on ex-ante forecasts. Techno-economic energy system models, used for ex-ante prediction, have the
key disadvantage they do not use recent historical prices to benchmark their price estimators. As a result, they may
struggle to explain random short-term variations compared to econometric models that "learn from the past". However,
energy system models possess an advantage in that they are based on established economic theory and replicate the

Watermeyer et al. Page 4 of 38

 A hybrid model for day ahead electricity price forecasting

workings of markets. As such, they are able to predict prices independently of past data and are less prone to structural
changes in the market and other similar factors. In line with that, we did not find techno-economic energy system
model applications to forecasts for ex-ante predictions of short-term electricity prices. However, there are such uses
in ex-ante predictions of long-term electricity prices, where random short-term variations are less inherent [see, e.g.,
Muesgens, 2020, Green and Vasilakos, 2010, Lamont, 2008]. In addition, technical-economic energy system models
have been employed in bodies of literature that extend beyond price estimates (e.g., to determine the value of demand
response [Misconel et al., 2021, Kirchem et al., 2020], to identify an optimal transmission-expansion plan [van der
Weijde and Hobbs, 2012], to support decision-making at the municipal level [Scheller and Bruckner, 2019], to analyse
the effect of power-to-gas [Lynch et al., 2019], to evaluate policy instruments to reduce CO2 emissions [Sgarciu et al.,
2023]). Additionally, Plaga and Bertsch [2023] thoroughly examines how energy system models can account for climate
uncertainty. A comprehensive overview of energy system modelling can be found in Ventosa et al. [2005].
In recent years, hybrid methods have garnered significant attention in electricity price forecasting. Hybrid models
are those that combine two or more distinct methods. They aim to use the combined strengths of the employed
methods while mitigating their individual weaknesses to achieve better overall results. Many hybrid methods have
been developed that combine a wide variety of methods. Aggarwal and Tripathi [2017], for example, present a hybrid
approach that uses a wavelet transform, a time-series time-delay neural network and an error-predicting algorithm to
predict day-ahead electricity prices in the ISO New England market. Chang et al. [2019] combine an Adam-optimised
LSTM neural network to generate electricity prices with a wavelet transform to decompose an electricity price time
series into several series of electricity prices. A combination of an empirical wavelet transform, a support vector
regression, a bi-directional LSTM and a Bayesian optimisation is proposed by Cheng et al. [2019]. Nazar et al. [2018]
apply a three-stage hybrid model to the DK2 area of Nordpool and the Spanish power market. The first stage features a
wavelet and Kalman machines to decompose price data into different frequency components. The second stage uses an
adaptive neuro-fuzzy inference system (ANFIS) to forecast price frequency components. In the third stage, the output
of the second stage is fed into the ANFIS to boost forecasting accuracy. A wavelet transform and an ARMA are paired
with a kernel-based extreme learning machine by Yang et al. [2017], and with a radial basis function neural network
by Olamaee et al. [2016]. Zhang et al. [2020] propose a hybrid model based on variational mode decomposition, self-
adaptive particle swarm optimisation, SARIMA and a deep belief network for short-term electricity price forecasting.
Most of the hybrid models mentioned above use statistical and deep learning methods. However, a few applications
also combine a techno-economic energy system model with another approach. For example, de Marcos et al. [2019]
detail a short-term hybrid electricity price forecasting model for the Iberian market that combines a techno-economic
cost-generation optimisation model with an ANN. Gonzalez et al. [2012] propose two hybrid approaches based on a
techno-economic electricity market model. Focusing on the day-ahead market in the UK, they combine this model type
separately with two other models: a linear autoregressive model with exogenous data on price drivers (ARX model)
and a nonlinear logistic regression model with a smooth transition (LSTR model), which is a regime change in times of
structural change. Their results support the idea of incorporating fundamental information for better price forecasting.
Particularly in highly volatile periods, the nonlinear hybrid model achieves better results. In Möbius et al. [2023], our
previous study, we introduced a techno-economic market model tailored to the day-ahead market and combined it with
a stochastic model to enhance day-ahead load forecast accuracy in the estimation of day-ahead electricity prices. We
highlighted the positive effects of better load forecasts on the day-ahead price estimators generated with an energy
system model. However, this approach merely represents a first step; it does not fully realise the potential of a hybrid
model, which seamlessly integrates the strengths of both the techno-economic and stochastic models.
The literature on electricity price forecasting mostly focuses on developing point forecasting methods for the day-
ahead market. However, in recent years, there has been a growing interest in probabilistic forecasting methods [Hong
et al., 2020]. The Global Energy Forecasting Competition (GEFCom2014) [see Hong et al., 2016] served as a catalyst
for this trend, and many studies have been published on this topic in the time since. Nowotarski and Weron [2018]
provide a comprehensive overview of the different approaches used in this field. A hybrid model combining point and
probabilistic forecasting in four steps was developed by Maciejowska and Nowotarski [2016] for the GEFCom2014.
Common approaches to probabilistic electricity price forecasting include using time-series models, such as ARIMA,
GARCH and exponential smoothing (ETS) [e.g., Weron and Misiorek, 2008] and using deep learning models.
Bootstrapping is widely used in combination with deep learning approaches [e.g., Chen et al., 2012, Wan et al.,
2014, Rafiei et al., 2017, Khosravi et al., 2013]. On top of deep learning, Zhao et al. [2008] use a support vector

Watermeyer et al. Page 5 of 38

 A hybrid model for day ahead electricity price forecasting

machine (SVM) to estimate prediction intervals and density forecasts, and Zhou et al. [2006] use an extended ARIMA
model to do the same. An econometric model for probabilistic forecasting is proposed by Panagiotelis and Smith
[2008]. Manner et al. [2019] use vine-copula models to forecast quantiles for a vector of day-ahead electricity prices
from interconnected electricity markets, while Grothe et al. [2023] propose an approach based on copula techniques
that entails generating multivariate probabilistic forecasts by modelling cross-hour dependencies. Considering these
dependencies in probabilistic forecasts is uncommon, in contrast to point forecasts. However, including them in the
methodology for generating probabilistic price forecasts can enhance forecast accuracy.
Historical simulation and distribution-based prediction intervals are other popular approaches. Historical simulation
estimates risk and generates prediction intervals in the simulation of multiple scenarios using historical data; it then uses
the results to estimate the probability of different outcomes [e.g., Weron and Misiorek, 2008, Nowotarski and Weron,
2015]. Distribution-based prediction intervals are calculated based on the distribution of historical data [e.g., Misiorek
et al., 2006, Zhao et al., 2008, Dudek, 2016, Maciejowska et al., 2016, Panagiotelis and Smith, 2008]. A theoretical
introduction to the generation of prediction intervals based on distribution and historical simulation is provided by
Weron [2006].
Quantile regression averaging (QRA) is a method that has risen in prominence recently in probabilistic electricity
price forecasting. It combines predictions from multiple quantile regression models, each of which is trained to predict
a different quantile of the response variable. This method was first formally introduced by Nowotarski and Weron
[2015] and has since continued to be applied and developed further due to its accuracy [e.g., Maciejowska et al., 2016,
Nowotarski and Weron, 2014, Marcjasz et al., 2020, Uniejewski et al., 2019, Uniejewski and Weron, 2021].
Despite the rising prominence of probabilistic forecasts in various models, there is still a general lack of approaches
that combine probabilistic forecasting with techno-economic energy system models. This paper aims to fill this gap in
the literature. By adapting and developing an energy system model specifically for the short-term electricity market
and combining this model with common stochastic models through multiple steps, we can leverage the strengths of
both models and open up the field of short-term electricity price forecasting for energy system models. Having already
highlighted the positive effects of combining a stochastic model (for better load forecasts) with an energy system model
(for the day-ahead market, developed by Möbius et al. [2023]), these building blocks are included in the hybrid model.
We demonstrate that a multi-layer hybrid model makes point and probabilistic price forecasting with techno-economic
and stochastic models possible.

3. Data
In our study, we develop a hybrid model that integrates stochastic modelling approaches and energy system optimisation
to forecast wholesale electricity prices. Notably, this energy system optimisation requires multiple inputs. Table 1
provides an overview of the necessary input data. In this section, we provide more details on how the data is obtained
and applied.
Although our modelling approach can be applied to many markets, our empirical exercise focuses exclusively on the
German spot market. However, the high level of integration among European electricity markets and the resulting
interdependencies require a comprehensive representation of these markets, particularly during the energy system
optimisation step. Figure 2 shows the geographical scope of the collected data and the interconnection among European
markets. We consider the bidding zones of most of the EU’s 27 member states1 as well as Norway, Switzerland and
the United Kingdom.2 Unless stated otherwise, the collected data is from 2016 to 2020.
Electricity demand is represented by hourly values for the system’s electrical load, of which both a day-ahead forecast
and the actual values are published by the respective transmission system operators (TSOs) and provided by ENTSO-
E Transparency Platform [2021a]. The collected load data for the Germany-Luxembourg bidding zone represents
2015–2020. In energy system models, electricity demand is usually considered volatile and inflexible in the short
1 Bulgaria, Cyprus, Greece, Iceland, Ireland, Malta and Romania are omitted.
2 Note that we aggregate the bidding zones of Spain and Portugal to a single ‘Iberian’ market and the bidding zones of Lithuania, Estonia
and Latvia to a single ‘Baltic’ market. Additionally, note that we consider the distinct bidding zones within countries. However, we aggregate the
following zones: in Norway, zones NO1–NO5; in Sweden, zones SE1–SE3; and in Italy, all zones but IT-North.

Watermeyer et al. Page 6 of 38

 A hybrid model for day ahead electricity price forecasting

Table 1
Overview of required data
Parameter Source
CO2 prices Sandbag [2020]
Control power procurement Regelleistung.net [2018]
Efficiency of generation capacities Schröder et al. [2013],
Open Power System Data [2020a]
Efficiency losses at partial load Schröder et al. [2013]
Electricity demand (load) ENTSO-E Transparency Platform [2021a]
Energy-power factor (for storages) own assumption: 9
Fuel prices Destatis Statistisches Bundesamt [2021],
(Lignite, nuclear, coal, gas, oil) ENTSO-E [2018], ENTSO-E [2018],
EEX [2021]
Generation and storage capacity BNetzA [2021], UBA [2020], EBC [2021],
ENTSO-E Transparency Platform [2021b],
Open Power System Data [2020b]
Generation by CHP units European Commission [2021]
Historic electricity generation ENTSO-E Transparency Platform [2021c]
Load shedding costs own assumption: 3,000 C/MWh
Minimum output levels Schröder et al. [2013]
NTCs ENTSO-E Transparency Platform [2021d],
JAO Joint Allocation Office [2021]
Variable O&M costs Schröder et al. [2013]
Power plant outages ENTSO-E Transparency Platform [2021e]
RES feed-in ENTSO-E Transparency Platform [2021f]
RES curtailment costs own assumption: 20 C/MWh
Start-up costs Schröder et al. [2013]
Seasonal availability of hydro power ENTSO-E Transparency Platform [2021c]
Temperature (daily mean) Open Power System Data [2020a]
Water value ENTSO-E Transparency Platform [2021c],
ENTSO-E Transparency Platform [2021g]

term. However, there is typically an option to shed load amid supply scarcity. In our application, we assume the cost
of load shedding to be 3,000 A
C/MWh, as this was the maximum bidding price prior to September 20223 .
The availability of intermittent renewable energy, namely onshore wind, offshore wind and photovoltaics (PV), depends
on meteorological conditions and varies from hour to hour. The feed-in data of these energy sources are provided as
hourly day-ahead forecasts by ENTSO-E Transparency Platform [2021f]. Despite weather dependency, renewable
energy electricity generation can still be intentionally curtailed. Acknowledging the various support schemes for
renewable generation in Germany and Europe, which prevent renewable sources from being shut down immediately
when negative prices occur, this study assumes a curtailment cost of 20 AC∕𝑀𝑊 ℎ𝑒𝑙 .
For conventional thermal generation, we distinguish between ten technologies and divide further by age if their
technical parameters (especially those that impact efficiency) have changed significantly over time. We use 30 capacity
clusters to group power plants based on technology and date of commissioning. We derive technology- and age-based
efficiencies from Open Power System Data [2020b] data and assign them to the corresponding capacity clusters.
Moreover, we assign the clusters minimum output levels and efficiency losses in part-load operations based on Schröder
et al. [2013]. Hence, supply that follows fluctuations in demand and renewables is incentivised to shut down due
to physical barriers (minimum output levels) and economic incentives (efficiency losses). The capacity, fuel type,
generation technology and date of commissioning for units in the German market are derived from BNetzA [2021],
UBA [2020] and Open Power System Data [2020b]. For the remaining markets considered in our study, we use data
from ENTSO-E Transparency Platform [2021b], Open Power System Data [2020b] and EBC [2021].
3 The maximum bidding price was increased to 4,000 A
C/MWh on 20 September 2022.

Watermeyer et al. Page 7 of 38

 A hybrid model for day ahead electricity price forecasting

Figure 2: Geographical scope of the energy system model.

Power plant efficiency and the costs of fuel and CO2 emissions form the variable generation costs of conventional
thermal technologies. For fuel costs, we apply daily gas prices that are provided by EEX [2021], monthly coal prices
taken from Destatis Statistisches Bundesamt [2021] and monthly oil prices from Destatis Statistisches Bundesamt
[2021]. Fuel costs for nuclear and lignite are derived from ENTSO-E [2018] and are assumed not to vary over the time
horizon of our study. Prices for CO2 certificates are taken as weekly data from Sandbag [2020].
Due to the time and additional fuel used by power plants to heat up during a start-up process, fuel and CO2 prices also
impact the cost of starting up a power plant. Further data regarding start-ups (e.g., secondary fuel usage, depreciation)
are derived from Schröder et al. [2013].
Electricity generation relies on both the installed capacity and technical availability of power plants. As a result,
we consider all scheduled and unscheduled outages that were known before the closure of the day-ahead market.
Information on hourly outages is obtained from ENTSO-E Transparency Platform [2021e].
In most electricity markets, combined heat and power (CHP) plants are used, where electricity generation and heat
supply are interconnected and reliant on each other. To account for this relationship, we apply a must-run condition
to all CHP units to ensure operation at a minimum output level, as defined by the heat demand. These output levels
are established in two steps. First, we calculate an hourly heat-demand factor consisting of temperature-dependent
(spatial heating) and temperature-independent (warm water and process heat) components. The temperature-driven
heat demand is calculated using heating degree days derived by mean temperature data obtained from Open Power
System Data [2020a], while the temperature-independent heat demand is obtained from hourly and daily consumption
patterns provided by Hellwig [2003]. Second, we allocate annual electricity generation volumes from CHP plants
to each hour of the year based on the hourly heat-demand factor. The data on annual technology-specific electricity
generation from CHP units is sourced from European Commission [2021].
Control power is essential for system operators to ensure frequency stability at all times. The day-ahead market is
affected by the market for control power provision, meaning that the capacities reserved for control power cannot be
placed on the day-ahead market. The amount of control power to be procured is an average of the tender results taken
from Regelleistung.net [2018].

Watermeyer et al. Page 8 of 38

 A hybrid model for day ahead electricity price forecasting

In addition to conventional thermal technologies and intermittent renewables, we consider waste, biomass, energy
storage, hydro-reservoirs and run-of-river hydroelectricity. Since the operation of both waste and biomass has largely
been historically constant (compare with ENTSO-E Transparency Platform [2021c]), we implement both technologies
as base-load.
The energy storage units are divided into high capacity-to-energy ratios and low capacity-to-energy ratios. Storage
units with high capacity-to-energy ratios actively charge and discharge. We exclusively consider a subset of pumped
storage plants (PSPs) in this group. The overall turbine capacity of these PSPs is detailed by ENTSO-E Transparency
Platform [2021b], and the efficiency of a storage cycle is around 75 % [Schröder et al., 2013]. For these PSPs, we
assume a capacity-to-energy ratio of 1/9 [see DENA, 2010]. This means that the plant can generate electricity at full
load for nine hours before its storage is emptied. Storage units with low capacity-to-energy ratios comprise long-term
PSPs as well as hydro-reservoirs. They are assigned a variable generation cost in the model (i.e., an opportunity cost
for water consumption). Using historical electricity prices from ENTSO-E Transparency Platform [2021g] and the
observed generation and pumping activities in the respective hour from ENTSO-E Transparency Platform [2021c],
we construct a step-wise merit order for long-term PSPs and hydro-reservoirs. Run-of-river and high-capacity-to-
energy PSPs are assigned a monthly availability factor derived from the historical generation data from ENTSO-E
Transparency Platform [2021c]. We assume that 70 % of pump storage capacity belongs to high-capacity-to-energy
PSPs, while the remaining 30 % of pump storage capacity belongs to low-capacity-to-energy PSPs.
European electricity markets are highly interconnected through cross-border transmission capacities. Thus, the German
electricity market is also integrated into the European electricity system, with a total interconnector capacity of 27 GW,
representing more than 30 % of the German peak load.4 Annual aggregated exports, accounting for approximately
13 % of German consumption in 2019, and imports, making up around 7 % of consumption in the same year, are both
significant. In our energy system model, net transfer capacities (NTCs) constrain transmission between market zones.
We implement hourly day-ahead forecasts for NTCs that are made available by ENTSO-E Transparency Platform
[2021d] and JAO Joint Allocation Office [2021].
While parameterised data sets interest numerous stakeholders, replicating these is a tremendous effort [see, e.g.,
Schröder et al., 2013, Kunz et al., 2017]; thus, our input data can be found in the supplementary materials.

4. Methodology
The implementation of the entire hybrid model is open-source; all data we could make public are available on GitHub at
the following link: https://github.com/ProKoMoProject/A-hybrid-model-for-day-ahead-electricity-price-forecasting-
combining-fundamental-and-stochastic-mod.
In this section, we present the components of the hybrid model, starting with the methodologies for the data pre-
processing and parameter density forecast steps in Sections 4.1 and 4.2. Next, we introduce the dispatch market model
used to generate the first price estimators in the energy system optimisation step in Section 4.3. Finally, we detail the
model for the stochastic post-processing step in Section 4.4.

4.1. Stochastic Data Pre-Processing Step

Modelling the electricity market with a dispatch model requires an understanding of the several various fundamental
variables, including demand (represented by load), which is a crucial element. The TSOs provide the actual load and
a day-ahead load forecast, which has the potential to be improved, as demonstrated by, for example, Maciejowska
et al. [2021] and Möbius et al. [2023]. To enhance this forecast, we use a stochastic model for data pre-processing.
Additionally, we model a two-day-ahead load forecast to capture power plant start-ups and shutdowns based on the
load of the second following day, using the TSOs’ day-ahead load forecast as a starting point.

Day-Ahead Load Forecast

We use the approach initially presented in Möbius et al. [2023] to improve load forecasting. Thus, we occasionally refer
to it for detailed specifications and analyses. We propose a purely endogenous time-series approach: a model for the
4 Note that the availability of the interconnectors depends on various factors (e.g., congestion within a market zone).

Watermeyer et al. Page 9 of 38

 A hybrid model for day ahead electricity price forecasting

TSOs’ load forecast error 𝜀𝑡 that depends only on past values of the forecast error itself and, in turn, on the TSOs’ load
forecast 𝑙̂𝑡 . The forecast error is the difference between the actual load data 𝑙𝑡 and the TSOs’ load forecast 𝑙̂𝑡 . Designing
a model for the error and forecasting it enables us to improve the TSOs’ load prediction. The resulting load prediction
∗
𝑙̂𝑡 at time 𝑡 is given by the following equation:
∗
𝑙̂𝑡 = 𝑙̂𝑡 + 𝜀̂𝑡 , (1)

where 𝜀̂𝑡 is our forecasted TSOs’ load prediction error. Thus, 𝑙̂∗ is an improved load forecast in which we adjust the
original forecast for the predictable structure of its error. The sub-index 𝑡 denotes consecutive hours.
To model and forecast the forecast error 𝜀𝑡 , we use a decomposition model and decompose the error time series into
the sum of a seasonal component and a stochastic component (see Lütkepohl [2005], Hyndman and Athanasopoulos
[2021], Box et al. [2015] for comprehensive introductions to time series models):

𝜀𝑡 = 𝑆𝐶𝑡 + 𝑅𝐶𝑡 , (2)

where 𝑆𝐶𝑡 is a seasonal and 𝑅𝐶𝑡 is the remaining, stochastic component at time 𝑡.
Capturing a weekly season, the seasonal component 𝑆𝐶𝑡 for time 𝑡 is defined by Eq. (3) with 𝐻𝑆 ℎ,𝑤𝑑 being the average
of TSOs’ forecast errors for hour ℎ = 1, ..., 24 and day 𝑤𝑑 = 1 (Monday), . . . , 7 (Sunday) and 𝐻𝑜𝑊𝑡ℎ,𝑤𝑑 , describing
dummy variables to address the hour of the day and the day of the week:
24 ∑
∑ 7
𝑆𝐶𝑡 = 𝐻𝑜𝑊𝑡ℎ,𝑤𝑑 ⋅ 𝐻𝑆 ℎ,𝑤𝑑 , (3)
ℎ=1 𝑤𝑑=1

with
∑𝑡−ℎ−24
𝑠=𝑡−ℎ−𝑙𝑤 −23 𝜀𝑠 ⋅ 𝐻𝑜𝑊𝑠ℎ,𝑤𝑑
ℎ,𝑤𝑑
𝐻𝑆 ∶= ∑𝑡−ℎ−24 , (4)
ℎ,𝑤𝑑
𝑠=𝑡−ℎ−𝑙𝑤 −23 𝐻𝑜𝑊𝑠

{
ℎ,𝑤𝑑 1, if 𝑡 is the h-th hour of the wd-th day of the week,
𝐻𝑜𝑊𝑡 =
0, otherwise.

For the residual component 𝑅𝐶𝑡 = 𝜀𝑡 − 𝑆𝐶𝑡 of the time series, we propose an econometric SARMA (1, 1)x(1, 1)24
model given by the following equation:

𝑅𝐶𝑡 =𝜙0 + 𝜙1 ⋅ 𝑅𝐶𝑡−1 + 𝜙24 ⋅ 𝑅𝐶𝑡−24 − 𝜙1 𝜙24 ⋅ 𝑅𝐶𝑡−25

+ 𝜔1 ⋅ 𝜓𝑡−1 + 𝜔24 ⋅ 𝜓𝑡−24 + 𝜔1 𝜔24 ⋅ 𝜓𝑡−25 (5)
+ 𝜓𝑡 ,

where the innovations are assumed to be homoscedastic and normally distributed. This model contains an additional
24-hour seasonal component, making it stochastic, flexible and dependent on the values of past hours and days. In
contrast, 𝑆𝐶𝑡 describes a static seasonality.
The model is estimated over a calibration window of one year. Within the hybrid model, the window is constantly rolled
over by full days, with the forecast of the next day’s 24 hours being made recursively. If no actual data are available
due to time points in the future, we use forecast values for the model variables.

Two-Day-Ahead Load Forecast

Power plants make operational start-up and shut-down decisions based on the current day’s demand, the demand from
the day before and the expected demand on the next day. To account for this in the dispatch model, a forecast of load
consumption is needed two days in advance. We use the modelling and forecast of the current load 𝑙𝑡 (meaning the

Watermeyer et al. Page 10 of 38

 A hybrid model for day ahead electricity price forecasting

TSOs’ day-ahead load forecast 𝑙̂𝑡 ) as a starting point from which to forecast the day-ahead forecast for the second
following day, resulting in a two-day-ahead forecast 𝑙̂2𝐷𝐴 = 𝑙̂̂𝑡 .
𝑡
For the model, we propose an econometric SARMA (1, 1)x(1, 1)24 model with an additional exogenous variable, the
TSOs’ load forecast at lag 168:

𝑙̂̂𝑡 =𝜙0 + 𝜙1 ⋅ 𝑙̂𝑡−1 + 𝜙168 ⋅ 𝑙̂𝑡−168 + 𝜙24 ⋅ 𝑙̂𝑡−24 − 𝜙1 𝜙24 ⋅ 𝑙̂𝑡−25

(6)
+ 𝜔1 ⋅ 𝜓𝑡−1 + 𝜔24 ⋅ 𝜓𝑡−24 + 𝜔1 𝜔24 ⋅ 𝜓𝑡−25 .

The model’s innovations 𝜓𝑡 are assumed to be homoscedastic and normally distributed, meaning that 𝜓𝑡 ∼ 𝑁(0, 𝜎𝜀2 ).
The model features 24-hour seasonality. Since we also observe weekly seasonality, we include the TSOs’ load forecast
at the same hour one week earlier, 𝑙̂𝑡−168 , as a regressor.
We calibrate and estimate the model based on window length 𝑙𝑤 , which contains one year of historical data. The
estimated model is used to recursively (i.e., on an hourly basis) predict the values of each hour of the next day. Since we
rely on an autoregressive time-series model, day-ahead load forecasts from 168 hours to one hour before the predicted
hour are in the model as explanatory variables for the two-day-ahead prediction. However, this means that some values
are unavailable when the forecast is made. They are replaced with recursively forecasted values based on the most
recently available observations.
Given the increased uncertainty associated with forecasting two days ahead, our hybrid model incorporates a parameter
density forecast that considers various scenarios for the level and development of load, utilising the hourly two-day-
ahead load forecast as an input variable.

4.2. Parameter Density Forecast Step

To account for uncertainty in two-day-ahead load predictions, scenarios are calculated at the 5 % and 95 % quantiles
using QRA. It describes a method for determining quantiles of predictive cumulative distribution functions, which can
then be used to construct prediction intervals. A prediction interval is calculated using the (𝛼∕2)-th and (1 − 𝛼∕2)-
th quantile of the predictive cumulative distribution function, 𝛼 ∈ (0, 1), as the lower and upper bound of the
interval. QRA is based on quantile regression and aims to model quantiles of real-valued variables that depend on
explanatory variables [see, e.g., Koenker and Bassett, 1978]. It employs point predictions to explain the 𝑞-th quantile
of the conditional distribution of the observation, setting a fixed 𝑞 ∈ (0, 1). Here, quantile regression uses a vector of
regressors 𝑋𝑡 = [1, 𝑙̂𝑡2𝐷𝐴 ], including a value of one for the intercept and the two-day-ahead point prediction for the load
at given time 𝑡, to calculate the two-day-ahead load prediction in the 𝑞-th quantile (𝑄𝑙𝑡 (𝑞|⋅)) (conditional on additional
information):

𝑄𝑙𝑡 (𝑞|𝑋𝑡 ) = 𝑋𝑡 𝛽𝑞 (7)

Thereby, 𝛽𝑞 is a vector of parameters for the 𝑞-th quantile. QRA estimates 𝛽𝑞 and, thus, the 𝑞-th quantile by minimising
the pinball loss function of the respective 𝑞-th quantile, given by
[ ∑ ]
𝛽̂𝑞 = argmin (𝑞 − 1𝑙𝑡 <𝑋𝑡 𝛽 )(𝑙𝑡 − 𝑋𝑡 𝛽) , (8)
𝛽 𝑡∶𝑙𝑡 ≥𝑋𝑡 𝛽

where 1 is the indicator function [see, e.g., Nowotarski and Weron, 2015, 2018].
We use the two-day-ahead load prediction 𝑙̂𝑡2𝐷𝐴 and the corresponding load from a one-year historical period to estimate
the unknown parameter 𝛽𝑞 . After calculating the 5 % and 95 % quantiles, the scenarios cover 90 % of possible load
forecast values with 𝛼 = 0.9. With the parameter density forecast and the point forecast of expected value, this approach
provides three possible scenarios for the two-day-ahead load: an expected scenario, described with the two-day-ahead
load forecast, a low scenario described with the hourly estimated 5 % quantile, and a high scenario described with
the hourly estimated 95 % quantile. Motivated by optimal integration rules in Grothe [2013] (and setting 𝜅 = 2, as
recommended in that work), we weight the expected value with 2∕3 and each quantile with 1∕6 to include the scenarios
in the em.power dispatch model described below.

Watermeyer et al. Page 11 of 38

 A hybrid model for day ahead electricity price forecasting

4.3. Energy System Optimisation Step

This section presents the energy system model em.power dispatch, which generates wholesale day-ahead price
estimators in the hybrid model’s energy system optimisation step. The model considers a detailed representation
of the key techno-economic aspects of an integrated European electricity sector, including transmission restrictions
between markets, electricity production by CHPs, energy storage and control power provision. For all considered
market zones, our model determines the optimal dispatch decisions for various generation and storage technologies, the
most effective use of cross-border transmission capacities and the short-run marginal system costs5 , which determine
the price estimator for the day-ahead market in hourly resolution.
Since our research focuses on day-ahead price forecasts, the energy system model is developed to reflect the level and
quality of information available to market participants on the day before delivery. The model is formulated as a linear
optimisation problem minimising total system costs. Ensuring the linear formulation of a highly complex system, we
form capacity clusters, parameterising them as described in Section 3. Within each capacity cluster, capacity units can
be started up, and electricity can be produced in marginal increments [see Muesgens, 2006b]. This approach has two
key advantages. First, it reduces computational requirements. Second, the problem is differentiable at each point, and
the dual variable of the demand constraint can be interpreted as a wholesale market price estimator. Additionally, the
accuracy of modelling large energy systems remains reasonably high for our purpose [Muesgens and Neuhoff, 2006].
We implement the resultant imperfect forecasts with two model features. First, we implement a rolling window
approach that repeatedly solves three days (𝑑 ∈ 𝐷 = {𝑑, 𝑑 + 1, 𝑑 + 2}), as shown in Figure 3. In this setting, the
24 hours of the target day are represented by the second day of the horizon (𝑑 + 1). This follows the EPEX spot
market organisation, in which 24 hourly day-ahead prices are determined at 12 p.m. on the day before delivery (𝑑). In
addition to the target day d+1, we include the day before (𝑑) and the day after (𝑑 + 2). By considering three days in our
rolling window, we reduce the problems of starting and ending values, particularly those stemming from power plant
start-ups and pump storage plants. Second, we account for the increased uncertainty of the two-day-ahead estimate
of key parameters. While parameters for the day 𝑑 are fully known, and forecasts for 𝑑 + 1 are made available by
the European network transmission system operators for electricity (Entso-e), the realisation of key input parameters
exhibits higher uncertainty in 𝑑 + 2. Therefore, we implement probabilistic forecast intervals only in 𝑑 + 2, as shown in
Figure 3. All other days are provided through one scenario. The resulting stochastic rolling window is then repeatedly
solved using only data available to market participants when they need to make their decisions. In each model run, we
extract the information for the 24 hours of our ‘target day’ – the day ahead.

Figure 3: Illustration of the rolling window.

The optimisation problem is repeatedly solved each day. For the reader’s convenience, we provide a nomenclature in
the appendix. Note that all endogenous variables are written in upper case, and all exogenous parameters are written
in lower case.
The objective function in Eq. (9) minimises total system costs (𝑇 𝐶) consisting of the expected value of all operational
costs (𝑂𝐶) across all three days of a rolling window 𝑑 ∗ . Our empirical exercise considers three scenarios 𝑠 to be equally
likely to appear.

min 𝑇 𝐶 = 𝔼𝑠 [𝑂𝐶(𝑠)] (9)

5 Technically, the dual variable of demand constraint derives the short-run marginal system cost, also called the ‘shadow price’.

Watermeyer et al. Page 12 of 38

 A hybrid model for day ahead electricity price forecasting

The operational costs in Eq. (10) contain all of the short-term costs that generation units face. We include costs at
𝐹 𝐿 ), additional costs for units that operate at partial load (𝑣𝑐 𝑀𝐿 − 𝑣𝑐 𝐹 𝐿 ) and start-up
full-load operation (𝑣𝑐𝑖,𝑛,𝑑,ℎ 𝑖,𝑛,𝑑,ℎ 𝑖,𝑛,𝑑,ℎ
costs (𝑠𝑐𝑖,𝑛,𝑑,ℎ ). Note that we apply a linear unit commitment formulation and that all units must produce at least
a certain minimum output level (see Eqs. (12) – (16)). Additionally, we account for load-shedding costs (𝑣𝑜𝑙𝑙) and
penalty payments for curtailing renewables (𝑐𝑢𝑟𝑡𝑐), as discussed in Section 3.

∑ ∑
𝐹𝐿
𝑂𝐶𝑠 = 𝐺𝑖,𝑛,𝑑,ℎ,𝑠 ⋅ 𝑣𝑐𝑖,𝑛,𝑑,ℎ + 𝑆𝑈𝑖,𝑛,𝑑,ℎ,𝑠 ⋅ 𝑠𝑐𝑖,𝑛,𝑑,ℎ
𝑖,𝑛,𝑑,ℎ 𝑖,𝑛,𝑑,ℎ
∑
𝑜𝑛 𝑀𝐿 𝐹𝐿
+ (𝑃𝑖,𝑛,𝑑,ℎ,𝑠 − 𝐺𝑖,𝑛,𝑑,ℎ,𝑠 ) ⋅ (𝑣𝑐𝑖,𝑛,𝑑,ℎ − 𝑣𝑐𝑖,𝑛,𝑑,ℎ ) ⋅ 𝑔𝑖𝑚𝑖𝑛 ∕(1 − 𝑔𝑖𝑚𝑖𝑛 )
𝑖,𝑛,𝑑,ℎ
∑ ∑
+ 𝐺𝑠𝑡𝑙,𝑛,𝑑,ℎ,𝑠 ⋅ 𝑤𝑣𝑠𝑡𝑙,𝑛,𝑑,ℎ − 𝐶𝐿𝑠𝑡𝑙,𝑛,𝑑,ℎ,𝑠 ⋅ 𝑤𝑣𝑠𝑡𝑙,𝑛,𝑑,ℎ (10)
𝑠𝑡𝑙,𝑛,𝑑,ℎ 𝑠𝑡𝑙,𝑛,𝑑,ℎ
∑ ∑
+ 𝐺ℎ𝑟,𝑛,𝑑,ℎ,𝑠 ⋅ 𝑤𝑣ℎ𝑟,𝑛,𝑑,ℎ + 𝑆𝐻𝐸𝐷𝑛,𝑑,ℎ,𝑠 ⋅ 𝑣𝑜𝑙𝑙
ℎ𝑟,𝑛,𝑑,ℎ 𝑛,𝑑,ℎ
∑
+ 𝐶𝑈 𝑅𝑇𝑟𝑒𝑠,𝑛,𝑑,ℎ,𝑠 ⋅ 𝑐𝑢𝑟𝑡𝑐
𝑟𝑒𝑠,𝑛,𝑑,ℎ

As we apply our model with a rolling window, each model run considers three days 𝑑 with 24 hours each day, meaning
72 hours per daily model run. Modelling an additional day before and after the target day seems appropriate for storage
units operated on a daily cycle. However, storage units (both PSPs and seasonal storage units without pumps) have a
storage cycle longer than three days. Therefore, PSPs are divided into low-capacity-to-energy storage, which operates
storage cycles longer than three days, and high-capacity-to-energy storage operating one or more storage cycles within
a three-day horizon. The dispatch of the latter is determined endogenously. Low-capacity-to-energy PSPs are assigned
a water value (𝑤𝑣𝑠𝑡𝑙,𝑛,𝑑,ℎ ) that is implemented as a variable cost factor for electricity generation (𝐺𝑠𝑡𝑙,𝑛,𝑑,ℎ,𝑠 ) and
electricity consumption (𝐶𝐿𝑠𝑡𝑙,𝑛,𝑑,ℎ,𝑠 ).
Moving beyond pumped storage plants, hydro reservoirs have a natural water feed-in and do not have pumps installed.
However, the water budget for electricity generation is limited by seasonal inflow volumes. We also apply a water value
(𝑤𝑣ℎ𝑟,𝑛,𝑑,ℎ ) to account for the opportunity costs of electricity generation from hydro reservoirs (𝐺ℎ𝑟,𝑛,𝑑,ℎ,𝑠 ).
Market clearing is ensured by Eq. (11). For all 72 hours of the given rolling window, demand (𝑙𝑛,𝑑,ℎ,𝑠 ) must equal
the sum of generation (𝐺𝑖,𝑛,𝑑,ℎ,𝑠 ), load shedding (𝑆𝐻𝐸𝐷𝑛,𝑑,ℎ,𝑠 ) and electricity imports (𝐹 𝐿𝑂𝑊𝑛𝑛,𝑛,𝑑,ℎ,𝑠 ) minus the
electricity consumption of mid-term energy storage (𝐶𝑀𝑠𝑡𝑚,𝑛,𝑑,ℎ,𝑠 ) and long-term energy storage (𝐶𝐿𝑠𝑡𝑙,𝑛,𝑑,ℎ,𝑠 ) as well
as electricity exports (𝐹 𝐿𝑂𝑊𝑛,𝑛𝑛,𝑑,ℎ,𝑠 ). The dual variable of the demand constraint Eq. (11) is used as an hourly
day-ahead wholesale electricity price estimator.
∑ ∑ ∑
𝑙𝑛,𝑑,ℎ,𝑠 = 𝐺𝑖,𝑛,𝑑,ℎ,𝑠 − 𝐶𝑀𝑠𝑡𝑚,𝑛,𝑑,ℎ,𝑠 − 𝐶𝐿𝑠𝑡𝑙,𝑛,𝑑,ℎ,𝑠
𝑖 𝑠𝑡𝑚⊂𝐼 𝑠𝑡𝑙⊂𝐼
∑
+ (𝐹 𝐿𝑂𝑊𝑛𝑛,𝑛,𝑑,ℎ,𝑠 − 𝐹 𝐿𝑂𝑊𝑛,𝑛𝑛,𝑑,ℎ,𝑠 ) + 𝑆𝐻𝐸𝐷𝑛,𝑑,ℎ,𝑠 (11)
𝑛𝑛
∀𝑛, 𝑛𝑛 ∈ 𝑁, 𝑑 ∈ 𝐷, ℎ ∈ 𝐻, 𝑠 ∈ 𝑆

Note that we apply the improved point forecast for the load created in Section 4.1 to Germany in d and d+1, as we
focus on price predictions for the German market. For all other markets, we implement the original Entso-e forecasts.
For d+2 in Germany, we apply the probabilistic load forecast presented in Section 4.2. For the remaining markets, we
use the actual realisation of the previous week as a more naive estimator.
Electricity generation by a capacity cluster is limited by upper and lower bounds. The upper bound is formalised
𝑜𝑛
in Eq. (12). It ensures that electricity generation does not exceed the running capacity (𝑃𝑖,𝑛,𝑑,ℎ,𝑠 ) in the cluster. The
potential to generate electricity by running capacity is further limited by the reserve for positive control power provision
(𝑃 𝐶𝑅𝑖,𝑛,𝑏𝑝,𝑠 and 𝑆𝐶𝑅𝑝𝑜𝑠𝑖,𝑛,𝑏𝑠,𝑠
). The lower bound is presented in Eq. (13). It states that running capacities must operate
at a minimum power level, including the capacity reserved for negative control power provision (𝑃 𝐶𝑅𝑖,𝑛,𝑏𝑝 and

Watermeyer et al. Page 13 of 38

 A hybrid model for day ahead electricity price forecasting

𝑆𝐶𝑅𝑛𝑒𝑔𝑖,𝑛,𝑏𝑠,𝑠
). Note that primary control power provision (𝑃 𝐶𝑅𝑖,𝑛,𝑏𝑝,𝑠 ) in Germany is symmetrical (i.e., a unit must
provide both positive and negative primary control power). Different positive and negative control power products were
introduced for secondary control power. We do not include minute reserve requirements in the model for two reasons.
First, fast-reacting units (e.g., hydro, open-cycle gas turbines) can be started up to provide positive minute reserve
without being dispatched. Second, both positive and negative reserves can be provided by multiple market players
other than the power plants included in the model (e.g., demand flexibility, P2X units, emergency power generators).
The hours that belong to bidding blocks are mapped for primary control power by 𝑏𝑝 and for secondary control power
by 𝑏𝑠.

𝑜𝑛
𝐺𝑖,𝑛,𝑑,ℎ,𝑠 ≤ 𝑃𝑖,𝑛,𝑑,ℎ,𝑠 − 𝑃 𝐶𝑅𝑖,𝑛,𝑏𝑝|ℎ∈𝑏𝑝,𝑠 − 𝑆𝐶𝑅𝑝𝑜𝑠
𝑖,𝑛,𝑏𝑠|ℎ∈𝑏𝑠,𝑠
(12)
∀𝑏𝑝 ∈ 𝐵𝑃 , 𝑏𝑠 ∈ 𝐵𝑆, 𝑖 ∈ 𝐼, 𝑛 ∈ 𝑁, 𝑑 ∈ 𝐷, ℎ ∈ 𝐻, 𝑠 ∈ 𝑆

𝑜𝑛
𝑃𝑖,𝑛,𝑑,ℎ,𝑠 ⋅ 𝑔𝑖𝑚𝑖𝑛 + 𝑃 𝐶𝑅𝑖,𝑛,𝑏𝑝|ℎ∈𝑏𝑝,𝑠 + 𝑆𝐶𝑅𝑛𝑒𝑔
𝑖,𝑛,𝑏𝑠|𝑡∈𝑏𝑠,𝑠
≤ 𝐺𝑖,𝑛,𝑑,ℎ,𝑠
(13)
∀𝑏𝑝 ∈ 𝐵𝑃 , 𝑏𝑠 ∈ 𝐵𝑆, 𝑖 ∈ 𝐼, 𝑛 ∈ 𝑁, 𝑑 ∈ 𝐷, ℎ ∈ 𝐻, 𝑠 ∈ 𝑆

The installed capacity limits the running capacity of a power system (𝑐𝑎𝑝𝑖,𝑛,𝑑,ℎ ) in combination with either the
availability factor (𝑎𝑓𝑖,𝑛,𝑑,ℎ ) or power plant outages (𝑜𝑢𝑡𝑖,𝑛,𝑑,ℎ ), as shown in Eq. (14). For thermal generation capacities,
we use hourly power plant outages. Renewables are provided with an hourly availability factor, while hydroelectric units
are provided with a monthly availability factor.

𝑜𝑛
𝑃𝑖,𝑛,𝑑,ℎ,𝑠 ≤ 𝑐𝑎𝑝𝑖,𝑛,𝑑,ℎ ⋅ 𝑎𝑓𝑖,𝑛 − 𝑜𝑢𝑡𝑖,𝑛,𝑑,ℎ ∀𝑖 ∈ 𝐼, 𝑛 ∈ 𝑁, ℎ ∈ 𝐻, 𝑑 ∈ 𝐷, 𝑠 ∈ 𝑆 (14)

Eq. (15) tracks start-up activities (𝑆𝑈𝑖,𝑛,𝑑,ℎ,𝑠 ) that increase running capacity from one hour to another. Due to the
non-negativity condition, start-ups are either positive or zero. Eq. (16) tracks start-up activities from the last hour of a
day to the first hour of the following day.

𝑜𝑛 𝑜𝑛
𝑃𝑖,𝑛,𝑑,ℎ,𝑠 − 𝑃𝑖,𝑛,𝑑,ℎ−1,𝑠 ≤ 𝑆𝑈𝑖,𝑛,𝑑,ℎ,𝑠 ∀𝑖 ∈ 𝐼, 𝑛 ∈ 𝑁, ℎ ∈ 𝐻, 𝑑 ∈ 𝐷, 𝑠 ∈ 𝑆 (15)

𝑜𝑛 𝑜𝑛
𝑃𝑖,𝑛,𝑑,ℎ𝑓 𝑖𝑟𝑠𝑡,𝑠 − 𝑃𝑖,𝑛,𝑑−1,ℎ𝑙𝑎𝑠𝑡,𝑠 ≤ 𝑆𝑈𝑖,𝑛,𝑑,ℎ𝑓 𝑖𝑟𝑠𝑡,𝑠 ∀𝑖 ∈ 𝐼, 𝑛 ∈ 𝑁, 𝑑 ∈ 𝐷, 𝑠 ∈ 𝑆 (16)

The difference between available feed-in from intermittent renewables and their actual generation defines the
curtailment of renewables, as shown in Eq. (17):

𝑐𝑎𝑝𝑟𝑒𝑠,𝑛,𝑑,ℎ ⋅ 𝑝𝑓𝑟𝑒𝑠,𝑛,𝑑,ℎ = 𝐺𝑟𝑒𝑠,𝑛,𝑑,ℎ,𝑠 + 𝐶𝑈 𝑅𝑇𝑟𝑒𝑠,𝑛,𝑑,ℎ,𝑠

(17)
∀𝑛 ∈ 𝑁, 𝑟𝑒𝑠 ∈ 𝐼, 𝑑 ∈ 𝐷, ℎ ∈ 𝐻, 𝑠 ∈ 𝑆.

Some power plants are active in both the heat market and the electricity market. Thus, the model implements a must-run
condition for such units on the electricity market, which varies over time (e.g., higher in the winter season due to space
heating). Depending on hourly heat demand, Eq. (18) states that the output of a combined heat and power unit is at
least equal to the electricity generation linked to the heat production (𝑐ℎ𝑝𝑖,𝑛,𝑑,ℎ ):

𝑐ℎ𝑝𝑖,𝑛,𝑑,ℎ ≤ 𝐺𝑖,𝑛,𝑑,ℎ,𝑠 ∀𝑖 ∈ 𝐼, 𝑛 ∈ 𝑁, 𝑑 ∈ 𝐷, ℎ ∈ 𝐻, 𝑠 ∈ 𝑆. (18)

Watermeyer et al. Page 14 of 38

 A hybrid model for day ahead electricity price forecasting

Eq. 19 constrains cross-border electricity transfer (𝐹 𝐿𝑂𝑊𝑛,𝑛𝑛,𝑑,ℎ,𝑠 ) via net transfer capacity (𝑛𝑡𝑐𝑛,𝑛𝑛,𝑑,ℎ ):

𝐹 𝐿𝑂𝑊𝑛,𝑛𝑛,𝑑,ℎ,𝑠 ≤ 𝑛𝑡𝑐𝑛,𝑛𝑛,𝑑,ℎ ∀𝑛, 𝑛𝑛 ∈ 𝑁, 𝑑 ∈ 𝐷, ℎ ∈ 𝐻, 𝑠 ∈ 𝑆. (19)

Eq. (20) describes the state of the storage level of mid-term storage units. A storage level decreases with electricity
𝑖𝑛
generation (𝐺𝑠𝑡𝑚,𝑛,𝑑,ℎ ) and increases with charging (𝑆𝑇𝑠𝑡𝑚,𝑛,𝑑,ℎ,𝑠 ). The efficiency of an entire storage cycle (𝜂𝑠𝑡𝑚 ) is
assigned to the charging process. Eq. (21) ensures the functionality of the storage mechanism between two days:

𝑆𝐿𝑠𝑡𝑚,𝑛,𝑑,ℎ,𝑠 = 𝑆𝐿𝑠𝑡𝑚,𝑛,𝑑,ℎ−1,𝑠 − 𝐺𝑠𝑡𝑚,𝑛,𝑑,ℎ,𝑠 + 𝐶𝑀𝑠𝑡𝑚,𝑛,𝑑,ℎ,𝑠 ⋅ 𝜂𝑠𝑡𝑚

(20)
∀𝑠𝑡𝑚 ∈ 𝐼, 𝑛 ∈ 𝑁, 𝑑 ∈ 𝐷, ℎ ∈ 𝐻, 𝑠 ∈ 𝑆,

𝑆𝐿𝑠𝑡𝑚,𝑛,𝑑,ℎ𝑓 𝑖𝑟𝑠𝑡,𝑠 =𝑆𝐿𝑠𝑡𝑚,𝑛,𝑑−1,ℎ𝑙𝑎𝑠𝑡,𝑠 − 𝐺𝑠𝑡𝑚,𝑛,𝑑,ℎ𝑓 𝑖𝑟𝑠𝑡,𝑠

+ 𝐶𝑀𝑠𝑡𝑚,𝑛,𝑑,ℎ𝑓 𝑖𝑟𝑠𝑡,𝑠 ⋅ 𝜂𝑠𝑡𝑚 (21)
∀𝑠𝑡𝑚 ∈ 𝐼, 𝑛 ∈ 𝑁, 𝑑 ∈ 𝐷, 𝑠 ∈ 𝑆.

The maximum energy storage capacity (𝑆𝐿𝑠𝑡𝑚,𝑛,𝑑,ℎ,𝑠 ) of storage units with high capacity-to-energy ratios is defined by
their maximum installed turbine capacity divided by their capacity-to-energy ratio (𝑐𝑒𝑟), as shown in Eq. (22):

𝑆𝐿𝑠𝑡𝑚,𝑛,𝑑,ℎ,𝑠 ≤ 𝑐𝑎𝑝𝑠𝑡𝑚,𝑛,𝑑,ℎ ⋅ 𝑐𝑒𝑟 ∀𝑠𝑡𝑚 ∈ 𝐼, 𝑛 ∈ 𝑁, 𝑑 ∈ 𝐷, ℎ ∈ 𝐻, 𝑠 ∈ 𝑆. (22)

Eq. (23) restricts both turbine capacity and pumping capacity, with pumping capacity assumed to be 10 % lower than
the turbine capacity:

𝐺𝑠𝑡𝑚,𝑛,𝑑,ℎ,𝑠 + 1.1 ⋅ 𝐶𝑀𝑠𝑡𝑚,𝑛,𝑑,ℎ,𝑠 ≤ 𝑐𝑎𝑝𝑠𝑡𝑚,𝑛,𝑑,ℎ

(23)
∀𝑠𝑡𝑚 ∈ 𝐼, 𝑛 ∈ 𝑁, 𝑑 ∈ 𝐷, ℎ ∈ 𝐻, 𝑠 ∈ 𝑆.

At the beginning and at the end of each model run, all storage units with high capacity-to-energy ratios must be filled
with at least 30 % of their energy level:

𝑆𝐿𝑠𝑡𝑚,𝑛,𝑑𝑓 𝑖𝑟𝑠𝑡,ℎ𝑓 𝑖𝑟𝑠𝑡,𝑠 = 0.3 ⋅ 𝑐𝑎𝑝𝑠𝑡𝑚,𝑛,𝑑,ℎ ∀𝑠𝑡𝑚 ∈ 𝐼, 𝑛 ∈ 𝑁, 𝑠 ∈ 𝑆, (24)

𝑆𝐿𝑠𝑡𝑚,𝑛,𝑑𝑙𝑎𝑠𝑡,ℎ𝑙𝑎𝑠𝑡,𝑠 = 0.3 ⋅ 𝑐𝑎𝑝𝑠𝑡𝑚,𝑛,𝑑𝑙𝑎𝑠𝑡,ℎ𝑙𝑎𝑠𝑡 ∀𝑠𝑡𝑚 ∈ 𝐼, 𝑛 ∈ 𝑁, 𝑠 ∈ 𝑆. (25)

Storage plants with low capacity-to-energy ratios are not subject to a storage mechanism. However, these units’
electricity generation and consumption are restricted to their installed capacity by:

𝐺𝑠𝑡𝑙,𝑛,𝑑,ℎ,𝑠 + 𝐶𝐿𝑠𝑡𝑙,𝑛,𝑑,ℎ,𝑠 ≤ 𝑐𝑎𝑝𝑠𝑡𝑙,𝑛,𝑑,ℎ ∀𝑛 ∈ 𝑁, 𝑠𝑡𝑙 ∈ 𝐼, 𝑑 ∈ 𝐷, ℎ ∈ 𝐻, 𝑠 ∈ 𝑆. (26)

Eqs. (27), (28) and (29) ensure the control power provision for primary, positive secondary and negative secondary
control power.

∑
𝑃 𝐶𝑅𝑖,𝑛,𝑏𝑝,𝑠 = 𝑝𝑟𝑛 ∀𝑏𝑝 ∈ 𝐵𝑃 , 𝑛 ∈ 𝑁, 𝑠 ∈ 𝑆 (27)
𝑖

Watermeyer et al. Page 15 of 38

 A hybrid model for day ahead electricity price forecasting

∑
𝑆𝐶𝑅𝑝𝑜𝑠
𝑖,𝑛,𝑏𝑠,𝑠
= 𝑠𝑟𝑝𝑜𝑠
𝑛 ∀𝑏𝑠 ∈ 𝐵𝑆, 𝑛 ∈ 𝑁, 𝑠 ∈ 𝑆 (28)
𝑖

∑
𝑆𝐶𝑅𝑛𝑒𝑔
𝑖,𝑛,𝑏𝑠,𝑠
= 𝑠𝑟𝑛𝑒𝑔
𝑛 ∀𝑏𝑠 ∈ 𝐵𝑆, 𝑛 ∈ 𝑁, 𝑠 ∈ 𝑆 (29)
𝑖

The non-negativity constraint ensures that the individual variables do not show negative values and is given by:

0 ≤𝐶𝐿𝑠𝑡𝑙,𝑛,𝑑,ℎ,𝑠 , 𝐶𝑀𝑠𝑡𝑚,𝑛,𝑑,ℎ,𝑠 , 𝐶𝑈 𝑅𝑇𝑟𝑒𝑠,𝑛,𝑑,ℎ,𝑠 ,

𝐺𝑖,𝑛,𝑑,ℎ,𝑠 , 𝐹 𝐿𝑂𝑊𝑛,𝑛𝑛,𝑑,ℎ,𝑠 , 𝑃𝑖,𝑛,𝑑,ℎ,𝑠 , 𝑃 𝐶𝑅𝑖,𝑛,𝑏𝑝,𝑠 , 𝑆𝐶𝑅𝑛𝑒𝑔
𝑖,𝑛,𝑏𝑠,𝑠
, 𝑆𝐶𝑅𝑝𝑜𝑠
𝑖,𝑛,𝑏𝑠,𝑠
,
(30)
𝑆𝐻𝐸𝐷𝑛,𝑑,ℎ,𝑠 , 𝑆𝐿𝑠𝑡𝑚,𝑛,𝑑,ℎ,𝑠 , 𝑆𝑈𝑖,𝑛,𝑑,ℎ,𝑠
∀𝑛, 𝑛𝑛 ∈ 𝑁, 𝑏𝑝 ∈ 𝐵𝑆, 𝑏𝑠 ∈ 𝐵𝑆, 𝑖 ∈ 𝐼, 𝑛 ∈ 𝑁, ℎ ∈ 𝐻, 𝑑 ∈ 𝐷, 𝑠 ∈ 𝑆.

4.4. Stochastic Data Post-Processing Step

In this step, we use a stochastic post-processing technique to refine the estimators produced by the techno-economic
energy system model. Specifically, we forecast the errors 𝜀𝑡 of the day-ahead price estimators 𝑃̂𝑡 obtained from the
energy system optimisation step, either by a time series based point forecast 𝜀̂𝑡 or by inferring the forecast errors
distribution functions to generate probabilistic day-ahead price predictions. We incorporate exogenous variables such
as renewable energy feed-in and weather, as well as lags of the forecast error itself into the time-series forecast of 𝜀𝑡 .
∗
We start with the improved point prediction 𝑃̂𝑡 at time 𝑡 which is given by the following equation:

∗
𝑃̂𝑡 = 𝑃̂𝑡 + 𝜀̂𝑡 , (31)

where 𝑃̂𝑡 is the price prediction from the last step, and 𝜀̂𝑡 is our model’s forecasted price prediction error. Thus, 𝑃̂ ∗
constitutes an improved price forecast in which we adjust the forecast from the last step for the stochastic but predictable
structure in its error.
We employ two model frameworks – univariate and multivariate – as this approach has been proven to be useful in
past research Ziel and Weron [2018]. In the univariate framework, we interpret the forecast error time series as one
high-frequency time series in an hourly resolution. In the multivariate framework, we split the time series into 24
individual time series, one for each hour, making them in a daily resolution.
For the post-processing setup, subindexes ℎ, 𝑑 will denote hours one through 24 of day 𝑑, with 𝑑 being consecutive
days. So, 𝑃1,1 , for instance, is the actual day-ahead price of the first hour of the first day of the considered period, and
𝜀5,432 is the error of the price estimator of the fifth hour of the 432nd day. This fits best because it enables us to observe
a realisation of 24 prices for the hours of the next day simultaneously for electricity prices. Please note that if ℎ − 1
were equal to or less than zero, we would need to shift one day backwards. Likewise, if ℎ + 1 were greater than 24, we
would need to shift one day forward.
We chose a standard econometric time-series model for the univariate framework. It consists of endogenous (i.e.,
autoregressive with moving average structures) and exogenous variables, all of which are integrated into a regression
model given by Eq. (32). To address several seasonal structures included in the time series of the prices’ forecast
errors 𝜀ℎ,𝑑 , we use the first and second observation backwards 𝜀ℎ−1,𝑑 , 𝜀ℎ−2,𝑑 as well as the first back error 𝜓ℎ−1,𝑑 of
the estimated model. Additionally, we use the observation one day before (daily structure) 𝜀ℎ,𝑑−1 and one week before
(weekly structure) 𝜀ℎ,𝑑−7 as endogenous explanatory variables. Considering daily effects, we include the minimum
and maximum forecast errors for the day before 𝜀𝑚𝑖𝑛,𝑑−1 , 𝜀𝑚𝑎𝑥,𝑑−1 . To account for the strong effects of forecast errors

Watermeyer et al. Page 16 of 38

 A hybrid model for day ahead electricity price forecasting

on public holidays, we use a dummy variable ℎ𝑜𝑙ℎ,𝑑 for public holidays as another factor. Additionally, we include an
hourly wind forecast 𝑋ℎ,𝑑 .

𝜀ℎ,𝑑 =𝜙0 + 𝜙1 ⋅ 𝜀ℎ−1,𝑑 + 𝜙2 ⋅ 𝜀ℎ−2,𝑑 + 𝜙3 ⋅ 𝜀ℎ,𝑑−1 + 𝜙4 ⋅ 𝜀ℎ,𝑑−7 + 𝜙5 ⋅ 𝜓ℎ−1,𝑑

+ 𝜔1 ⋅ 𝜀𝑚𝑖𝑛,𝑑−1 + 𝜔2 ⋅ 𝜀𝑚𝑎𝑥,𝑑−1 + 𝜔3 ⋅ ℎ𝑜𝑙ℎ,𝑑 + 𝜔4 ⋅ 𝑋ℎ,𝑑 (32)
+ 𝜓ℎ,𝑑

As in the univariate framework, we use the well-known time-series model ARX in the multivariate framework.
However, the autoregressive component refers to values of the same hour on previous days. The endogenous variables,
𝜀ℎ,𝑑−1 and 𝜀ℎ,𝑑−7 , are the forecast errors at the same hour one day prior and seven days prior, respectively. The
exogenous variables are the same as in the univariate framework: minimum and maximum forecast errors for the
day before, a dummy variable for public holidays and an hourly wind forecast. Thus, the model for the multivariate
framework is given by the following equation:

𝜀ℎ,𝑑 =𝜙0 + 𝜙1 ⋅ 𝜀ℎ,𝑑−1 + 𝜙2 ⋅ 𝜀ℎ,𝑑−7

+ 𝜔1 ⋅ 𝜀𝑚𝑖𝑛,𝑑−1 + 𝜔2 ⋅ 𝜀𝑚𝑎𝑥,𝑑−1 + 𝜔3 ⋅ 𝐷ℎ,𝑑 + 𝜔4 ⋅ 𝑋ℎ,𝑑 (33)
+ 𝜓ℎ,𝑑 ,

where 𝜙𝑖 , 𝜔𝑖 describes the coefficients that need to be estimated. The innovations 𝜓 are assumed to be homoscedastic
and normally distributed in both frameworks, meaning that 𝜓ℎ,𝑑 ∼ 𝑁(0, 𝜎𝜖2 ).
Since we rely on an autoregressive time-series model, we need day-ahead spot prices from the last hours before
prediction time as explanatory variables. In the multivariate framework, only one step into the future is forecasted
at a time due to the separate modelling of each hour. In the univariate framework, 24 values are forecast for the future,
with the hours of the next day predicted recursively (i.e., on an hour-by-hour basis). Unavailable variables are replaced
with recursively forecasted variables based on the most recent available observations.
In line with approaches previously shown to be effective in the literature [see, e.g., Marcjasz et al., 2018], we vary the
presented post-processing models by using different window lengths to estimate the model set-up and prevent random
choice. We determine the calibration window for 44, 48 and 52 weeks. By using three window lengths and two model
frameworks, we end up with six individual sub-models. These sub-models are used to predict the values of the hours
𝑢𝑣44 , 𝑃̂ 𝑢𝑣48 , 𝑃̂ 𝑢𝑣52 , 𝑃̂ 𝑚𝑣44 , 𝑃̂ 𝑚𝑣48 , 𝑃̂ 𝑚𝑣52 .
ℎ of the next day 𝑑, and we denote them by 𝑃̂ℎ,𝑑 ℎ,𝑑 ℎ,𝑑 ℎ,𝑑 ℎ,𝑑 ℎ,𝑑
∗
Our final improved point forecast 𝑃̂𝑡 is obtained by taking the arithmetic average of the six prices. Despite the potential
appeal of using seemingly more sophisticated methods, such as calculating optimal weights via linear regression based
on past data, these methods resulted in predictors with higher root mean squared errors (RMSE) and mean absolute
errors (MAE), even when we used rolling windows that look into the future. This is mainly due to the additional
estimation noise that is introduced when using such methods and may lead to inefficiencies as discussed, e.g., in the
context of financial literature in DeMiguel et al. [2009]. As a result, we stick with the simpler, yet more robust method
of averaging the six individual forecasts.
We now move on to generating probabilistic day-ahead price predictions. To achieve this, we use the six forecasts
generated by the individual sub-models to estimate the cumulative distribution function 𝐹𝑃ℎ,𝑑 of the day-ahead prices.
The estimated function 𝐹̂𝑃 serves as our probabilistic forecast for the price of the next day. We represent the
ℎ,𝑑
distribution 𝐹̂𝑃ℎ,𝑑 in terms of its quantiles. Specifically, we employ quantile regression to model the conditional 𝑞-
th quantile of the cumulative distribution function of the day-ahead prices, where 𝑞 is a value between 0 and 1. This
modelling is accomplished by utilising the six individual point predictions and the following equation:

𝑄𝑃ℎ,𝑑 (𝑞|𝑋ℎ,𝑑 ) = 𝑋ℎ,𝑑 𝛽𝑞 , (34)

Watermeyer et al. Page 17 of 38

 A hybrid model for day ahead electricity price forecasting

where 𝑋ℎ,𝑑 = [1, 𝑃̂ℎ,𝑑

𝑢𝑣44 , 𝑃̂ 𝑢𝑣48 , 𝑃̂ 𝑢𝑣52 , 𝑃̂ 𝑚𝑣44 , 𝑃̂ 𝑚𝑣48 , 𝑃̂ 𝑚𝑣52 ] is the vector of regressors containing a value of 1 for
ℎ,𝑑 ℎ,𝑑 ℎ,𝑑 ℎ,𝑑 ℎ,𝑑
the intercept and the six individual point predictions for the day-ahead price at time ℎ, 𝑑. 𝛽𝑞 is again estimated by
minimising the pinball loss function. To determine the predictive distribution, we forecast multiple quantiles of the
distribution.
The coefficients of the regressors are estimated by a calibration window of one year with a distinction made between
peak and off-peak hours. Peak hours are defined as those between 8 a.m. and 8 p.m. from Monday to Friday, while
off-peak hours are all remaining hours. This distinction is made because peak hours are characterised by high demand
for electricity and, therefore, often exhibit higher day-ahead prices. For each quantile 𝑞, we estimate two parameter
vectors 𝛽𝑞 (see Eq. (8)). Estimating one parameter vector 𝛽𝑞 for all day hours proved to be less accurate. Additional
information on this matter can be provided upon request.
In summary, our hybrid model includes the following steps. To predict the next day, we first enhance the TSO’s day-
ahead load forecast and predict the two-day-ahead load forecast in the stochastic data pre-processing step. We then
calculate the two-day-ahead load scenarios in the parameter density forecast step and include them in the em.power
dispatch in the energy system optimisation step to generate the first price estimators for the day-ahead spot market.
Finally, in the final stochastic post-processing step, we improve these price estimators with stochastic methods and
conduct probabilistic price forecasts. This sequence is repeated continuously, day by day, for all points in time in
our observation period. The hybrid model’s rolling window approach means that we always use the most up-to-date
available data.

5. Hybrid Model Results

In this section, we present the electricity price forecasts of our hybrid model for Germany from January 2016 to
December 20206 . Since the day-ahead market is organised in an hourly resolution, our hybrid model calculates point
and probabilistic forecasts for each hour of the following day. As the central point of this paper, we present the
overall results of the model (i.e., the point and probabilistic price predictions) and qualitatively detail their place in the
literature.
We start by comparing the point forecasts of our hybrid model to those in the literature. With an annual average
RMSE of 7.38 A C/MWh and MAE of 4.60 A C/MWh over the five years from 2016 to 2020, our model aligns well
with previous studies. An expert model developed by Ziel and Weron [2018] forecasted electricity prices with an
overall MAE of 5.01 A C/MWh for 2012 to 2016. Using an autoregressive model with exogenous variables, Maciejowska
et al. [2021] achieved an RMSE of 8.43 A C/MWh and MAE of 5.92 A C/MWh for 2016 to 2019. For the same period,
Qussous et al. [2022] obtained an RMSE of 11.21 A C/MWh and MAE of 7.89 A C/MWh, presenting an agent-based power
market-simulation model with rule-based bidding strategies and, thus, a non-equilibrium-oriented techno-economic
market model aimed at reproducing day-ahead electricity prices. Since they provided extensive information on the
error measures, we can use these models for a detailed comparison. However, as the time periods in these studies are
overlapping with but not identical to our observation period, we also perform a detailed comparison with the LEAR
model developed by Lago et al. [2021]. The LEAR model’s code is freely available, so we can extend it with data
from the years up to 2020. Furthermore, its day-ahead price forecasts are among the most precise in the literature,
and its authors are leading scholars in the field of price forecasting. Thus, we can perform a year-by-year comparison
with our results. Table 2 presents the forecast accuracy of the hybrid model developed in this paper, the agent-based
market simulation model and the LEAR model, showing the annual RMSE and MAE. It can be seen that the agent-
based model has the highest error, which can be attributed to the general difficulties of techno-economic models in
making short-term forecasts (the results of the em.power dispatch model without further post-processing steps are
compiled in Section 6.2 and point in the same direction). In contrast, the LEAR model and the proposed hybrid model
presented here exhibit similar error measures without larger gaps. Although the LEAR model more often takes the lead,
its advantage is limited, as it is a statistical model primarily designed for generating price forecasts. Comparatively,
the hybrid model’s forecast encompasses the entire market state represented by the energy system model, including
6 More precisely, prices cover the German-Austria-Luxembourg bidding zone from January 2016 to September 2018 and the German-

Luxembourg bidding zone from October 2018 to December 2020.

Watermeyer et al. Page 18 of 38

 A hybrid model for day ahead electricity price forecasting

Table 2
RMSE and MAE of day-ahead electricity price forecast through the presented hybrid model and the LEAR model in
[C/MWh]
RMSE MAE
Hybrid Agent LEAR Hybrid Agent LEAR
All years 7.38 11.21 7.24 4.60 7.89 4.38
2016 5.82 8.83 5.55 3.48 6.54 3.30
2017 8.79 13.01 7.91 5.25 9.44 4.56
2018 7.28 11.69 6.96 5.07 8.88 4.84
2019 7.05 10.91 7.95 4.43 6.69 4.53
2020 7.63 7.54 4.77 4.65

information on additional parameters of interest to market participants (e.g., CO2 emissions, international electricity
exchange, and power plant utilization).7
A more detailed evaluation of the prediction errors of the hybrid model is provided in Figures 4 and 5, which show the
RMSE for different criteria, such as base, peak and off-peak hours, and the hours of actual day-ahead price quantiles,
respectively. Again, peak hours are those between 8 a.m. and 8 p.m. from Monday to Friday; off-peak hours are the
remaining hours. Base hours describe all hours of a day, regardless of the day of the week or hour of the day.

Figure 4: RMSE for base, peak and off-peak hours.

While the errors of 7.47 AC/MWh in peak hours and 7.33 A C/MWh in off-peak hours seem quite similar over the whole
period, Figure 5 provides more insight. We separate the predicted hours into five groups presenting the hours for each
confidence interval of realised day-ahead prices in 20 % steps. Therefore, we calculate and evaluate the RMSE for each
group. The figure suggests a relationship between the error in price forecasts and the price level. Throughout the years,
the highest RMSE has consistently been identified in the hours with the lowest 20 % and highest 20 % of prices. Based
on economic theory, this finding can be explained by start-up costs and their impact on hourly prices. Assuming perfect
foresight, Kuntz and Muesgens [2007] have shown that start-up costs are added to fuel costs exclusively during the
hour of highest demand in a cycle because additional capacity must be started-up for that hour, which is not needed in
any other hour. During the hour of lowest demand, start-up costs are deducted from variable production costs because
7 Note that while this work focuses exclusively on prices, this model can also be informative about other factors, as discussed in the literature

review.

Watermeyer et al. Page 19 of 38

 A hybrid model for day ahead electricity price forecasting

power plants save costs on re-starts when allowed to continue operations throughout that hour. In contrast, start-up
costs do not influence prices during any other hour of a load cycle. The em.power dispatch model follows this economic
theory when determining wholesale electricity prices based on the shadow prices of the demand constraint. However,
in reality, bidders on the day-ahead market face uncertainties with regard to which hour has the highest and lowest
residual demand and what magnitude start-up costs have for that day. While uncertainty is always present, its impact is
likely higher when start-up costs need to be considered in addition to fuel costs and thus increase price volatility around
the highest and lowest price hours. Therefore, negative and positive price peaks are harder to capture and forecast than
intermediate price levels. Note that this increased uncertainty in these market conditions affects all point forecasting
models. Corresponding figures for the LEAR model, with a very similar pattern, are available in Appendix, Figures
A.1 and A.2.

Figure 5: RMSE for hours at different day-ahead price quantiles.

We now turn to the analysis of our probabilistic forecasts. We represent the forecast distribution of the day-ahead prices
for each forecast hour with quantile estimates in 5 % steps. The quantile estimation for different quantile levels then
enables the calculation of forecast intervals, which determine the probability of the day-ahead price being within a
certain range.
The first thing to check about probabilistic forecasts is whether they are well calibrated [see, e.g., Gneiting et al.,
2007], which means whether forecasted probabilities and observed frequencies coincide. Therefore, Figure 6 shows
the actual frequency of all hours in which the day-ahead prices are above or below the forecast quantile limits. A
forecast is considered calibrated if the predicted probabilities match the observed frequencies of the target variable
over time. Thus, a fully calibrated forecast in a laboratory environment would correspond to a uniform distribution.
The histogram below shows slightly increasing frequencies towards the outside but forms a good approximation of a
uniform distribution generated by randomly drawn numbers. This serves as a quality check for our hybrid model and
ensures that the predicted probabilities reflect the true probabilities of the day-ahead prices.
Since peak hours often have higher day-ahead prices than off-peak hours, our hybrid model predicts quantiles with two
QRA estimations for each quantile, one for peak hours and one for off-peak hours. Thus, the probabilistic forecasts
should be assessed for calibration separately in two disjunctive sub-sets. Figure 7 illustrates the calibration of the
quantiles for peak hours on the left and off-peak hours on the right. During off-peak hours, there is a higher frequency
in the outermost 5 % quantiles on both sides (i.e., at high and low price levels). During peak hours, more hours exceed
the quantile values, especially on the right side of the median. For example, 6.3 % of the hours exceed the 95 % quantile
value. With a higher frequency in the low 5 % quantile, the lowest prices can be attributed to the price-reducing influence
of high electricity generation from renewable energy sources. This is consistent with the proportion of peak hours that

Watermeyer et al. Page 20 of 38

 A hybrid model for day ahead electricity price forecasting

Figure 6: Number of actual prices included in the quantiles estimated in the forecast period from 2017 to 2020.

show high demand for electricity but also a high feed-in of renewable energies due to high solar radiation or wind
speeds and, thus, a high supply.

Figure 7: Number of actual prices included in the quantiles estimated in the 2017–2020 forecast period, separated for
peak (left) and off-peak (right) hours.

The calibration analysis of the predicted quantiles validates the calculated probabilistic forecasts. In the following, we
evaluate the hourly distribution of the width of the forecast intervals to obtain more precise and quantified information
about market uncertainty and the forecast accuracy that depends on it. Furthermore, we use probabilistic forecasts to
calculate the probability of negative prices. Thus, we address the economic interest in including probabilistic forecasts
in trading strategies and power plant deployment planning.
During peak hours, the uncertainty and fluctuations of the day-ahead prices are higher than during off-peak hours. This
can be explained by a steeper merit order at high-load levels, the occurrence of start-ups during peak hours and the
uncertainty of generation from renewable sources (especially PV plants, which produce more during peak hours). Our
hybrid model captures this uncertainty with its probabilistic forecasts. Figure 8 presents box plots that visualise the
distribution of the spread between the estimated 95 % quantile and the estimated 5 % quantile for each hour of the day
(the width of the 90 % prediction interval. From 8 a.m. to 8 p.m. (daytime), the median spread is wider and outliers are
higher than they are at night, reflecting the higher uncertainty and wider range of prices. Although the wider spread of

Watermeyer et al. Page 21 of 38

 A hybrid model for day ahead electricity price forecasting

the prediction interval covering 90 % of the potential prices shows that these hours are more difficult to forecast than
the night hours, the hybrid model can forecast them as accurately as it can the night hours. Figure 4 shows comparable
point forecast accuracy for peak and off-peak hours over the entire period. Additionally, the probabilistic forecasts
effectively mirror the market’s uncertainty.

Figure 8: Box plot of the 90 % prediction interval for each hour of the day.

Figure 9 shows the average width of the 90 % prediction interval for each hour of the week, illustrating how easy or
difficult it is to forecast the day-ahead electricity price for each individual hour of the week. A wider spread indicates
greater uncertainty and less predictability, as the price may take on a wider range of potential values. As previously
stated, daytime hours generally exhibit wider intervals than nighttime hours. However, more nuanced patterns are also
discernible.
On weekdays (Monday to Friday), the spread of the prediction interval follows a distinct daily course featuring three
concave curves with (local) maxima at 3 a.m., 8 a.m. and 4 p.m. The interval is relatively narrow from midnight to 7
a.m. From 7 a.m. to 8 a.m., the interval increases abruptly, posing a major challenge for the forecasting of electricity
prices, especially between 4 p.m. and 6 p.m. On weekends (Saturday and Sunday), the interval width is more evenly
distributed across all hours but remains high overall and still exhibits a notable decrease between 7 p.m. and 9 p.m.
Therefore, predicting day-ahead electricity prices requires careful consideration of the hour of the day and the day of
the week on top of all other factors that can affect market dynamics.
Figure 10 shows an example of how the probabilistic forecasts account for uncertainty. It shows the course of lower
and upper limits in combination with the forecast’s expected price (the price’s point prediction). The presented limits
correspond to the estimated 5 % and 95 % quantiles, respectively, and indicate the range of possible outcomes. More
precisely, actual day-ahead electricity price has a 90 % chance of lying within this range. Using the density forecast and
predicted bounds, we can make informed risk estimates and probability statements, which can be useful for decision-
making amid uncertainty.
Figure 11 illustrates the probabilities of negative prices for each hour of the week, as estimated by the density forecast.
Notably, negative prices can occur when, for example, increases in generation by renewable energy sources or a large
share of must-run capacities (e.g., CHP) force conventional power plant units to shut down, leading to start-up costs
when these capacities are needed again. The merit order effect of renewable energies and the regulation of renewable
energies, in particular payments for production besides the wholesale electricity market, can exacerbate this situation.

Watermeyer et al. Page 22 of 38

 A hybrid model for day ahead electricity price forecasting

Figure 9: Average spread between the predicted 5 % and 95 % quantiles for the hours of the week.

Figure 10: Day-ahead price forecast, lower-bound forecast (5 % quantile) and upper-bound forecast (95 % quantile).

Figure 11 reveals that the probability of negative prices is highest on Sundays and public holidays, reaching 10 to
15 % in some hours. This reflects the increased occurrence of negative prices in these hours in the actual day-ahead
prices. This is due to the fact that electricity demand is relatively low during these periods, increasing the likelihood
of negative price events. The probability of negative prices is also relatively high (above 5 %) in the early hours of
Monday, followed by a slightly higher probability of negative prices in the early hours of the other days of the week.
This pattern reflects the low-load behaviour of electricity demand in the early hours combined with the typically high
wind feed-in during the night and early morning, which can lead to price drops. On Sundays and holidays, the daytime
generation from PV and continuous wind feed-in can increase the likelihood of negative prices due to reduced demand.
The probability values for weekdays (Monday to Friday) are similar in their level and pattern, as these days have
comparable fundamental parameters. With the probabilities determined by the density forecasts, bidding strategies,

Watermeyer et al. Page 23 of 38

 A hybrid model for day ahead electricity price forecasting

Figure 11: Probability of negative day-ahead prices.

portfolios hedges and power plant usage can be optimised. The risks of negative prices or large price spreads can be
incorporated into strategies through the probability statements.
In summary, the results show the high-quality point and probabilistic price forecasts of the hybrid model, which also
offers two additional advantages over previous approaches. The integration of a techno-economic energy system model
allows for a deeper understanding of (energy) markets, and the use of stochastic approaches enables probabilistic price
forecasts to quantify uncertainty in these markets and provide valuable information on the probability of negative prices
and other market phenomena, helping market participants to make informed decisions and mitigate risks.

6. Analysis of Individual Model Steps

The hybrid model presented in this paper is multi-layered. In this section, we offer some insights into the individual
components (steps) of the hybrid model. Section 6.1 shows the results of the day-ahead load forecasts. Section 6.2
provides an evaluation of the price estimators that follow the energy system optimisation step. Finally, section 6.3
points out the impact of the six individual price estimators’ error forecasts.

6.1. Load Forecasts

This section presents the results of the load forecasts, where we compare the RMSE and MAE of TSOs’ load forecasts
with the improved load forecasts in Table 3, showing a significant improvement of 19 % in RMSE and 27 % in MAE.
We refer to Möbius et al. [2023] for a more detailed analysis.

6.2. Price Estimators After the Energy System Optimisation Step

In optimising the energy system model, our hybrid model generates price estimators derived from the dual variable of
demand constraint (see Section 4.3). Computing the dual variable, the model optimises the energy sector’s complex
technical and economic interdependencies. In addition to the costs of electricity generation, the model considers
unit commitment decisions, such as the start-up and shutdown of generation units, storage operation, limitations on
electricity transport to and from neighbouring countries, heat-supply requirements and the provision of control power.
All of these techno-economic interdependencies determine the resulting electricity prices and are essential in our
model, as they signal both high price peaks and low or even negative electricity prices observable on the market.

Watermeyer et al. Page 24 of 38

 A hybrid model for day ahead electricity price forecasting

Table 3
RMSE and MAE for the original TSO day-ahead load forecast (TSO) and the improved day-ahead load forecast (Impr.),
given in [𝑀𝑊 ℎ].
year all 2016 2017 2018 2019 2020
TSO 2,335.49 2,596.48 1,802.61 2,360.51 2,454.70 2,383.23
RMSE Impr. 1,881.64 1,872.44 1,483.45 2,043.87 1,665.76 1,859.29
% Improvement 19.43 27.89 17.71 13.41 32.14 21.98
TSO 1,797.05 2,028.44 1,396.45 1,726.67 1,951.00 1,881.84
MAE Impr. 1,320.83 1,302.17 1,106.12 1,372.55 1,283.17 1,405.71
% Improvement 26.50 35.80 20.79 20.51 34.23 25.30

Table 4
Descriptive statistics of the error of energy system optimisation step in [C/MWh].
all years 2016 2017 2018 2019 2020
mean -2.17 -2.23 -3.06 0.02 -0.73 -4.84
median -1.75 -2.24 -2.75 0.71 -0.23 -3.94
minimum -143.85 -143.85 -102.45 -81.45 -80.01 -79.05
maximum 105.73 50.62 105.73 55.19 42.87 86.76
5%-quantile -15.30 -11.81 -17.77 -13.24 -12.78 -19.50
95%-quantile 9.60 7.47 10.68 11.40 9.75 6.05
Std. 9.25 7.26 11.62 8.61 8.05 9.31
RMSE 9.50 7.60 12.01 8.61 8.09 10.49
MAE 6.00 4.83 7.09 5.91 5.07 7.13

Table 4 presents the descriptive statistics of the price estimation errors as well as the RMSE and MAE of these price
estimators after the energy optimisation step. For all years, the RMSE is 9.50 A C/MWh, and the MAE is 6.00 A C/MWh.
The lowest errors can be observed in 2016, and the highest errors can be observed in 2017. In comparison to state-
of-the-art electricity price forecasting models in the literature, the error measurements are larger, and the errors still
show a structural behaviour. For most years, the error’s mean and median are both negative values, meaning that the
em.power dispatch calculates prices higher than the observed prices on the market. With a value of -15.30 A C/MWh for
the error’s 5 % quantile and a value of 9.60 A C/MWh for the error’s 95 % quantile, 90 % of the error values lie in this
interval.
Figure 12 plots the error structure of the entire period divided into the hours of a week. Negative values indicate
an average overestimation of the prices in the corresponding hour, while positive values indicate an average
underestimation of the prices in the corresponding hour. The figure shows that the errors tend to increase over the
weekend, while there is a distinct daily pattern observed (see also Figure 13).
The model tends to underestimate prices during the night while overestimating them in the morning and late
afternoon, which is a consistent pattern across all years from 2016 to 2019. While the error structure in 2020 shows
some differences, possibly attributed to the COVID-19 pandemic, the model’s RMSE per hour of the day does not
significantly differ from previous years. We also observe a midday increased error rate and an evening error peak
between 4 p.m. to 6 p.m., which is more pronounced in the 2020 data.
Note, however, that we are evaluating here the short-term price prediction of a techno-economic energy system model,
which we expect, due to its inability to learn from history, to map the general market situation at these time horizons,
but not to form very accurate price estimators. Therefore, the model performs reasonably well with predictable patterns
in the errors, which indicates a high potential for the data post-processing step. The effect of this step is presented in
the following chapter.
Figure 14 illustrates the correlation of the price forecast errors with several exogenous variables. Evidently, wind
generation correlates significantly with the price forecast errors from the energy optimisation step and explains 14 %
of the overall forecast error’s volatility.

Watermeyer et al. Page 25 of 38

 A hybrid model for day ahead electricity price forecasting

Figure 12: Mean price estimator errors for the hours of a week (including public holidays at hours 168–192) after the
energy system optimisation step in [A
C/MWh].

Table 5
Descriptive statistics of the error of the hybrid model in [C/MWh]
all years 2016 2017 2018 2019 2020
mean -0.09 0.30 -0.15 1.12 -0.73 -0.99
median -0.08 0.08 -0.23 1.06 -0.68 -0.50
minimum -133.44 -133.44 -81.29 -54.41 -79.27 -69.43
maximum 86.10 43.76 74.30 52.72 35.48 86.10
5%-quantile -9.58 -6.64 -10.94 -9.09 -9.36 -11.51
95%-quantile 10.07 8.48 11.21 11.70 9.08 8.24
Std. 7.38 5.82 8.79 7.20 7.01 7.56

6.3. Improvement of Price Estimators

The detected systematic deviations and seasonal patterns in the errors of the price estimators following the energy
system optimisation step can be captured and predicted by the stochastic model in the post-processing step in a way
that improves the price forecast’s RMSE for 2016–2020 by 22 % through the combination of both model classes. The
statistical data of the hybrid model’s forecast error show that the error is distributed nearly symmetrically with a mean
of -0.09 A C/MWh and a median of -0.08 A C/MWh. Post-processing centres the price forecast error to near zero. At
the same time, both the minimum and maximum errors generated decrease each year and over the entire period – as
does overall volatility. Thus, the post-processing model achieves a reduction of 20 % in the standard deviation of the
day-ahead price forecast, compared to the one of the price estimators of the energy system optimisation step.
The stochastic post-processing step consists of six individual sub-models, as described in Section 4.4. Each of these
sub-models improves the price estimators obtained after the energy system optimisation step, significantly increasing
the forecast quality, as measured in Table 6 by specifying the error measures RMSE and MAE. While the models
with multivariate frameworks have lower error measures than those with univariate frameworks over the entire period,
the price forecasts of the univariate models are qualitatively better than those of the multivariate models in 2019 and

Watermeyer et al. Page 26 of 38

 A hybrid model for day ahead electricity price forecasting

Figure 13: Mean price estimator errors for the hours of a day in each year after the energy system optimisation step in
[A
C/MWh].

Figure 14: Correlation of possible exogenous variables and the price estimation error.

2020. Figure 15 shows the hourly RMSE for each sub-model. Depending on the length of the calibration window,
the models of the univariate framework dominate those of the multivariate framework in the first five hours of a
day. From that point forward, the multivariate sub-models achieve lower error measures. The result aligns with the
expectations of the model specifications. In the univariate framework, the time series is captured as a high-frequency
data set, so the forecast includes data up to the last available hour. However, this also requires 24 future values to be
forecasted, heightening inaccuracy. With the multivariate framework, a reverse effect is observable due to the split into

Watermeyer et al. Page 27 of 38

 A hybrid model for day ahead electricity price forecasting

Table 6
Error measurements RMSE and MAE of the error time series in [C/MWh].
Initial Post-processing
ESM UV44w UV48w UV52w MV44w MV48w MV52w Combination
All years 9.50 7.57 7.59 7.59 7.51 7.50 7.51 7.38 22 %
2016 7.60 6.07 6.13 6.09 5.95 5.94 5.92 5.82 23 %
RMSE 2017 12.01 9.22 9.20 9.13 8.92 8.86 8.86 8.79 27 %
2018 8.61 7.49 7.47 7.54 7.40 7.38 7.40 7.28 15 %
2019 8.09 7.11 7.13 7.14 7.21 7.22 7.21 7.05 13 %
2020 10.49 7.65 7.69 7.71 7.77 7.79 7.85 7.63 27 %
All years 6.00 4.73 4.75 4.75 4.75 4.74 4.74 4.60 23 %
2016 4.83 3.69 3.71 3.69 3.64 3.62 3.61 3.48 28 %
MAE 2017 7.09 5.55 5.55 5.51 5.37 5.31 5.28 5.25 26 %
2018 5.91 5.20 5.18 5.21 5.24 5.23 5.25 5.07 14 %
2019 5.07 4.46 4.49 4.49 4.61 4.61 4.59 4.43 13 %
2020 7.13 4.77 4.80 4.83 4.91 4.94 4.99 4.77 33 %

24 individual time series, as this split entails even just short steps into the future needing to be forecasted. Still, the
most recent available information is only partially considered.

Figure 15: RMSE for univariate and multivariate post-processing sub-models for each hour of the day in [A
C/MWh].

The individual characteristics of the univariate and multivariate frameworks also explain the average errors for each
hour of the day (see Figure A.4) and each hour of the week (see Figure A.5). While the forecasts of the multivariate
sub-models do not indicate a clear daily pattern in the under- or overestimation of values, the univariate sub-models
show a similar daily structure as the price estimators following the energy system optimisation step: underestimation
during the hours of 4–8 a.m. and 2–4 p.m. and overestimation,e.g., during the hours of 17-18 p.m. and for the hours from
21 p.m. The average error for weekly hours displays a significantly higher overestimation and underestimation on public
holidays. Notably, however, this trend is evident across all forecasts of the univariate and multivariate sub-models.
Table 6 presents the error measures for the price forecast that stems from combining the forecasts of individual sub-
models. Thus, it depicts the final price forecast of the hybrid model. The RMSE and MAE of this price forecast are
lower than the error measures of the individual price forecasts over each individual year as well as the entire period.
Using the multivariate Diebold-Mariano test from the epftoolbox by Lago et al. [2021] to compare different model
forecasts via hypothesis testing and determine whether one forecast’s accuracy is significantly higher than that of the
others, we show that the combination of the sub-models is significantly better than all six individual sub-models. The
test results are shown in Figure 16 as a heatmap illustrating the p-values of the hypothesis that the forecast of the
model on the Y-axis is significantly more accurate than the forecast of the model on the X-axis. A p-value close to

Watermeyer et al. Page 28 of 38

 A hybrid model for day ahead electricity price forecasting

zero indicates that the model on the X-axis has a significantly higher forecast accuracy than the model on the Y-axis.
The test indicates that, for forecasting day-ahead spot prices, the post-processing model and each stochastic sub-model
are significantly better than the energy system model modified by stochastic pre-processing and interweaving, which
means the hybrid model after the third step.

Figure 16: P-value of Diebold-Mariano test for all parts of the stochastic post-processing step (output derived from
epftoolbox by Lago et al. [2021]).

To better attribute and understand the effect of the stochastic post-processing step, we also determine the improvement
in RMSE for each hour of the day and day of the week, as shown in Figure 17. Daytime and weekday structures, which
can be observed in the price estimators’ errors after the energy system optimisation step, are evident in the improvement.
This is due in large part to the seasonal components in the stochastic model and the affected autoregressive structures.
Notably, the most substantial improvements in terms of percentage are achieved at night. Additionally, weekends and
especially holidays see relatively high levels of improvement. In the price estimators, these are the hours and days
with the most significant mean error, meaning those with a sizeable mean error and, in turn, the most potential for
improvement. Enhancing day-ahead price forecasts by reducing this error is a key goal of combining techno-economic
energy system modelling and stochastic modelling.

Figure 17: Average RMSE improvement in day-ahead price forecasts for each hour of the day (left) and each day of
the week (right).

Watermeyer et al. Page 29 of 38

 A hybrid model for day ahead electricity price forecasting

7. Conclusion
This paper introduced a novel hybrid model for short-term electricity price forecasting, which combines stochastic
modelling with (fundamental) techno-economic energy system modelling. The model consists of four steps. First,
through stochastic data pre-processing, day-ahead load forecasts are significantly improved, and a load forecast for two
days ahead is created. Second, this load forecast is extended by quantile forecasts in three different probable scenarios
of hourly load consumption two days in advance using a parameter density forecast step. Third, the modelled quantities
and input data are fed into the em.power dispatch, a techno-economic market model adapted for predicting day-ahead
price estimators. Fourth, the errors of the price estimators from the energy optimisation step are reduced by stochastic
models through stochastic data post-processing and complemented by probabilistic forecasts.
The hybrid model presented in this paper combines the strengths of stochastic models (strength: trained with price
data) and energy system models (strength: insights based on economic theory extending beyond prices). It can be
parameterised with data from most energy markets worldwide, providing insights into a wide range of energy systems.
We demonstrated its performance using German day-ahead electricity market data from January 2016 to December
2020. The price forecasts from the model had an annual average RMSE of 7.38 A C/MWh and an annual average MAE
of 4.60 A
C/MWh. The hybrid model’s forecasting accuracy surpasses the majority of benchmarks in the literature and
matches the best statistical benchmark identified in the literature. This confirms that techno-economic energy system
models can provide valuable input for price forecasting, even for short-term forecasting.
We conducted an in-depth analysis of price forecast quality and identified a notable relationship between the error in
price forecasts and the price level. Over the years, the highest RMSE has consistently been found in the hours with the
lowest 20 % and the highest 20 % of prices. This result can be explained by start-up costs and their impact on hourly
prices, as discussed in Section 5, and naturally affects all forecast models.
Additionally, our model provides calibrated density forecasts alongside point forecasts. With these, it can, for example,
quantify the probability of negative price events. Our analysis revealed that some hours on Sundays and public holidays
have a probability of more than 10 % of becoming negative.
Future research should focus on determining the most suitable model type for cases with varying lead times in price
forecasts. As discussed in the first part of this paper, the relative strength of techno-economic models in handling
structural breaks suggests their greater suitability for long-term forecasts (e.g., years ahead), while stochastic models,
whether time-series-based or artificial intelligence-based, are better suited for short-term forecasts (e.g., intra-day).
Nevertheless, hybrid models like the one introduced in this paper for day-ahead forecasts or similar approaches that
combine the advantages of both model classes mentioned above could, if properly adapted, be suitable for performing
accurate forecasts with variable lead times. As a result, we think that it is crucial to further explore and develop hybrid
models to harness their potential for delivering accurate forecasts across various lead times, making them a valuable
addition to the existing forecasting methodologies.

Data Availability
The data and source code presented in this work are openly available through the following link:
https://github.com/ProKoMoProject/A-hybrid-model-for-day-ahead-electricity-price-forecasting-combining-
fundamental-and-stochastic-mod.

Acknowledgements
The work was supported by the German Federal Ministry of Economic Affairs and Climate Action through the research
project ProKoMo within the Systems Analysis Research Network of the 6th energy research program.

References
Rafał Weron. Electricity price forecasting: A review of the state-of-the-art with a look into the future. International Journal of Forecasting, 30(4):
1030–1081, 2014. ISSN 0169-2070. doi: 10.1016/j.ijforecast.2014.08.008.

Watermeyer et al. Page 30 of 38

 A hybrid model for day ahead electricity price forecasting

Jakub Nowotarski and Rafał Weron. Recent advances in electricity price forecasting: A review of probabilistic forecasting. Renewable and
Sustainable Energy Reviews, 81:1548–1568, 1 2018. ISSN 18790690. doi: 10.1016/j.rser.2017.05.234.
Tao Hong, Pierre Pinson, Shu Fan, Hamidreza Zareipour, Alberto Troccoli, and Rob J. Hyndman. Probabilistic Global Energy Forecasting
Competition 2014 and beyond. International Journal of Forecasting, 32:896–913, 2016. ISSN 01692070. doi: 10.1016/J.IJFORECAST.2016.
02.001.
Tao Hong, Pierre Pinson, Yi Wang, Rafał Weron, Dazhi Yang, and Hamidreza Zareipour. Energy Forecasting: A Review and Outlook. IEEE Open
Access Journal of Power and Energy, 7:376–388, 2020. doi: 10.1109/OAJPE.2020.3029979.
N. Amjady and M. Hemmati. Energy price forecasting - problems and proposals for such predictions. IEEE Power and Energy Magazine, 4(2):
20–29, 2006. doi: 10.1109/MPAE.2006.1597990.
Sanjeev Kumar Aggarwal, Lalit Mohan Saini, and Ashwani Kumar. Electricity price forecasting in deregulated markets: A review and evaluation.
International Journal of Electrical Power & Energy Systems, 31(1):13–22, 2009. ISSN 0142-0615. doi: 10.1016/j.ijepes.2008.09.003.
Rafał Weron and Florian Ziel. Electricity price forecasting. In Routledge Handbook of Energy Economics, pages 506–521. Routledge, 9 2019. doi:
10.4324/9781315459653-36.
Fotios Petropoulos, Daniele Apiletti, Vassilios Assimakopoulos, Mohamed Zied Babai, Devon K. Barrow, Souhaib Ben Taieb, Christoph Bergmeir,
Ricardo J. Bessa, Jakub Bijak, John E. Boylan, Jethro Browell, Claudio Carnevale, Jennifer L. Castle, Pasquale Cirillo, Michael P. Clements,
Clara Cordeiro, Fernando Luiz Cyrino Oliveira, Shari De Baets, Alexander Dokumentov, Joanne Ellison, Piotr Fiszeder, Philip Hans Franses,
David T. Frazier, Michael Gilliland, M. Sinan Gönül, Paul Goodwin, Luigi Grossi, Yael Grushka-Cockayne, Mariangela Guidolin, Massimo
Guidolin, Ulrich Gunter, Xiaojia Guo, Renato Guseo, Nigel Harvey, David F. Hendry, Ross Hollyman, Tim Januschowski, Jooyoung Jeon,
Victor Richmond R. Jose, Yanfei Kang, Anne B. Koehler, Stephan Kolassa, Nikolaos Kourentzes, Sonia Leva, Feng Li, Konstantia Litsiou,
Spyros Makridakis, Gael M. Martin, Andrew B. Martinez, Sheik Meeran, Theodore Modis, Konstantinos Nikolopoulos, Dilek Önkal, Alessia
Paccagnini, Anastasios Panagiotelis, Ioannis Panapakidis, Jose M. Pavía, Manuela Pedio, Diego J. Pedregal, Pierre Pinson, Patrícia Ramos,
David E. Rapach, J. James Reade, Bahman Rostami-Tabar, Michał Rubaszek, Georgios Sermpinis, Han Lin Shang, Evangelos Spiliotis,
Aris A. Syntetos, Priyanga Dilini Talagala, Thiyanga S. Talagala, Len Tashman, Dimitrios Thomakos, Thordis Thorarinsdottir, Ezio Todini,
Juan Ramón Trapero Arenas, Xiaoqian Wang, Robert L. Winkler, Alisa Yusupova, and Florian Ziel. Forecasting: theory and practice.
International Journal of Forecasting, 38(3):705–871, 2022. ISSN 0169-2070. doi: https://doi.org/10.1016/j.ijforecast.2021.11.001. URL
https://www.sciencedirect.com/science/article/pii/S0169207021001758.
Rick Steinert and Florian Ziel. Short- to mid-term day-ahead electricity price forecasting using futures. Energy Journal, 40:105–127, 2019. ISSN
01956574. doi: 10.5547/01956574.40.1.rste.
Jakub Nowotarski and Rafał Weron. On the importance of the long-term seasonal component in day-ahead electricity price forecasting. Energy
Economics, 57:228–235, 6 2016. ISSN 01409883. doi: 10.1016/J.ENECO.2016.05.009.
Florian Ziel and Rafał Weron. Day-ahead electricity price forecasting with high-dimensional structures: Univariate vs. multivariate modeling
frameworks. Energy Economics, 70:396–420, 2018. ISSN 0140-9883. doi: 10.1016/j.eneco.2017.12.016.
T. M. Christensen, A. S. Hurn, and K. A. Lindsay. Forecasting spikes in electricity prices. International Journal of Forecasting, 28:400–411, 4
2012. ISSN 01692070. doi: 10.1016/j.ijforecast.2011.02.019.
Michael Eichler, Oliver Grothe, Hans Manner, and Tuerk Dennis. Models for short-term forecasting of spike occurrences in Australian electricity
markets: A comparative study. The Journal of Energy Markets, 7(1):55–81, 3 2014. doi: 10.21314/JEM.2014.104.
Hans Manner, Dennis Türk, and Michael Eichler. Modeling and forecasting multivariate electricity price spikes. Energy Economics, 60:255–265,
2016. ISSN 0140-9883. doi: 10.1016/j.eneco.2016.10.006.
R.C. Garcia, J. Contreras, M. van Akkeren, and J.B.C. Garcia. A GARCH forecasting model to predict day-ahead electricity prices. IEEE
Transactions on Power Systems, 20(2):867–874, 2005. doi: 10.1109/TPWRS.2005.846044.
Emily Hickey, David G. Loomis, and Hassan Mohammadi. Forecasting hourly electricity prices using ARMAX–GARCH models: An application
to MISO hubs. Energy Economics, 34(1):307–315, 2012. ISSN 0140-9883. doi: 10.1016/j.eneco.2011.11.011.
Silvano Bordignon, Derek W. Bunn, Francesco Lisi, and Fany Nan. Combining day-ahead forecasts for british electricity prices. Energy Economics,
35:88–103, 1 2013. ISSN 01409883. doi: 10.1016/j.eneco.2011.12.001.
Peter Kosater and Karl Mosler. Can markov regime-switching models improve power-price forecasts? Evidence from German daily power prices.
Applied Energy, 83(9):943–958, 2006. ISSN 0306-2619. doi: 10.1016/j.apenergy.2005.10.007.
Michael Bierbrauer, Stefan Trück, and Rafał Weron. Modeling electricity prices with regime switching models. Lecture Notes in Computer Science,
3039:859–867, 2004. ISSN 16113349. doi: 10.1007/978-3-540-25944-2\_111.
Carlo Mari and Emiliano Mari. Deep learning based regime-switching models of energy commodity prices. Energy Systems, 2022. doi:
10.1007/s12667-022-00515-6.
Bartosz Uniejewski, Jakub Nowotarski, and Rafał Weron. Automated Variable Selection and Shrinkage for Day-Ahead Electricity Price Forecasting.
Energies, 9(8), 2016. ISSN 1996-1073. doi: 10.3390/en9080621.
Jesus Lago, Grzegorz Marcjasz, Bart De Schutter, and Rafał Weron. Forecasting day-ahead electricity prices: A review of state-of-the-art algorithms,
best practices and an open-access benchmark. Applied Energy, 293:116983, 2021. ISSN 0306-2619. doi: https://doi.org/10.1016/j.apenergy.
2021.116983. URL https://www.sciencedirect.com/science/article/pii/S0306261921004529.
Ioannis P. Panapakidis and Athanasios S. Dagoumas. Day-ahead electricity price forecasting via the application of artificial neural network based
models. Applied Energy, 172:132–151, 6 2016. ISSN 03062619. doi: 10.1016/j.apenergy.2016.03.089.
Jesus Lago, Fjo De Ridder, Peter Vrancx, and Bart De Schutter. Forecasting day-ahead electricity prices in Europe: The importance of considering
market integration. Applied Energy, 211:890–903, 2 2018. ISSN 03062619. doi: 10.1016/j.apenergy.2017.11.098.
Nima Amjady. Day-ahead price forecasting of electricity markets by a new fuzzy neural network. IEEE Transactions on Power Systems, 21(2):
887–896, 2006. doi: 10.1109/TPWRS.2006.873409.
Grzegorz Marcjasz, Bartosz Uniejewski, and Rafał Weron. On the importance of the long-term seasonal component in day-ahead electricity
price forecasting with NARX neural networks. International Journal of Forecasting, 35(4):1520–1532, 2019. ISSN 0169-2070. doi:

Watermeyer et al. Page 31 of 38

 A hybrid model for day ahead electricity price forecasting

10.1016/j.ijforecast.2017.11.009.
Malte Lehna, Fabian Scheller, and Helmut Herwartz. Forecasting day-ahead electricity prices: A comparison of time series and neural network
models taking external regressors into account. Energy Economics, 106:105742, 2022. ISSN 0140-9883. doi: 10.1016/j.eneco.2021.105742.
Ramiz Qussous, Nick Harder, and Anke Weidlich. Understanding Power Market Dynamics by Reflecting Market Interrelations and Flexibility-
Oriented Bidding Strategies. Energies, 15(2), 2022. ISSN 1996-1073. doi: 10.3390/en15020494.
Felix Muesgens. Quantifying Market Power in the German Wholesale Electricity Market Using a Dynamic Multi-Regional Dispatch Model. The
Journal of Industrial Economics, 54(4):471–498, 2006a. doi: 10.1111/j.1467-6451.2006.00297.x.
Severin Borenstein, James B. Bushnell, and Frank A. Wolak. Measuring Market Inefficiencies in California’s Restructured Wholesale Electricity
Market. American Economic Review, 92(5):1376–1405, December 2002. doi: 10.1257/000282802762024557.
Dogan Keles, Massimo Genoese, Dominik Möst, Sebastian Ortlieb, and Wolf Fichtner. A combined modeling approach for wind power feed-
in and electricity spot prices. Energy Policy, 59:213–225, 2013. ISSN 0301-4215. doi: https://doi.org/10.1016/j.enpol.2013.03.028. URL
https://www.sciencedirect.com/science/article/pii/S0301421513001821.
Lion Hirth. The market value of variable renewables: The effect of solar wind power variability on their relative price. Energy Economics, 38:
218–236, 2013. ISSN 0140-9883. doi: 10.1016/j.eneco.2013.02.004.
Frank Sensfuß, Mario Ragwitz, and Massimo Genoese. The merit-order effect: A detailed analysis of the price effect of renewable electricity
generation on spot market prices in Germany. Energy Policy, 36(8):3086–3094, 2008. ISSN 0301-4215. doi: 10.1016/j.enpol.2008.03.035.
Christian Pape, Simon Hagemann, and Christoph Weber. Are fundamentals enough? Explaining price variations in the German day-ahead and
intraday power market. Energy Economics, 54:376–387, 2 2016. ISSN 0140-9883. doi: 10.1016/J.ENECO.2015.12.013.
Felix Muesgens. Equilibrium prices and investment in electricity systems with co2-emission trading and high shares of renewable energies. Energy
Economics, 86:104107, 2020. ISSN 0140-9883. doi: 10.1016/j.eneco.2018.07.028.
Richard Green and Nicholas Vasilakos. Market behaviour with large amounts of intermittent generation. Energy Policy, 38(7):3211–3220, 2010.
ISSN 0301-4215. doi: 10.1016/j.enpol.2009.07.038. Large-scale wind power in electricity markets with Regular Papers.
Alan D. Lamont. Assessing the long-term system value of intermittent electric generation technologies. Energy Economics, 30(3):1208–1231,
2008. ISSN 0140-9883. doi: 10.1016/j.eneco.2007.02.007.
Steffi Misconel, Christoph Zöphel, and Dominik Möst. Assessing the value of demand response in a decarbonized energy system – A large-scale
model application. Applied Energy, 299:117326, 2021. ISSN 0306-2619. doi: 10.1016/j.apenergy.2021.117326.
Dana Kirchem, Muireann Á. Lynch, Valentin Bertsch, and Eoin Casey. Modelling demand response with process models and energy systems
models: Potential applications for wastewater treatment within the energy-water nexus. Applied Energy, 260:114321, 2020. ISSN 0306-2619.
doi: 10.1016/j.apenergy.2019.114321.
Adriaan Hendrik van der Weijde and Benjamin F. Hobbs. The economics of planning electricity transmission to accommodate renewables: Using
two-stage optimisation to evaluate flexibility and the cost of disregarding uncertainty. Energy Economics, 34(6):2089–2101, 2012. ISSN 0140-
9883. doi: 10.1016/j.eneco.2012.02.015.
Fabian Scheller and Thomas Bruckner. Energy system optimization at the municipal level: An analysis of modeling approaches and challenges.
Renewable and Sustainable Energy Reviews, 105:444–461, 2019. ISSN 1364-0321. doi: 10.1016/j.rser.2019.02.005.
Muireann Lynch, Mel T. Devine, and Valentin Bertsch. The role of power-to-gas in the future energy system: Market and portfolio effects. Energy,
185:1197–1209, 2019. ISSN 0360-5442. doi: 10.1016/j.energy.2019.07.089.
Smaranda Sgarciu, Daniel Scholz, and Felix Muesgens. How co2 prices accelerate decarbonisation – The case of coal-fired generation in Germany.
Energy Policy, 173:113375, 2023. ISSN 0301-4215. doi: 10.1016/j.enpol.2022.113375.
Leonie Sara Plaga and Valentin Bertsch. Methods for assessing climate uncertainty in energy system models — a systematic literature review.
Applied Energy, 331:120384, 2023. ISSN 0306-2619. doi: 10.1016/j.apenergy.2022.120384.
Mariano Ventosa, Álvaro Baíllo, Andrés Ramos, and Michel Rivier. Electricity market modeling trends. Energy Policy, 33(7):897–913, 2005. ISSN
0301-4215. doi: 10.1016/j.enpol.2003.10.013.
Abhinav Aggarwal and M M Tripathi. A novel hybrid approach using wavelet transform, time series time delay neural network, and error predicting
algorithm for day-ahead electricity price forecasting. In 2017 6th International Conference on Computer Applications In Electrical Engineering-
Recent Advances (CERA), pages 199–204, 2017. doi: 10.1109/CERA.2017.8343326.
Zihan Chang, Yang Zhang, and Wenbo Chen. Electricity price prediction based on hybrid model of adam optimized LSTM neural network and
wavelet transform. Energy, 187:115804, 2019. ISSN 0360-5442. doi: 10.1016/j.energy.2019.07.134.
Hangyang Cheng, Xiangwu Ding, Wuneng Zhou, and Renqiang Ding. A hybrid electricity price forecasting model with Bayesian optimization
for German energy exchange. International Journal of Electrical Power & Energy Systems, 110:653–666, 2019. ISSN 0142-0615. doi:
10.1016/j.ijepes.2019.03.056.
Mehrdad Setayesh Nazar, Ashkan Eslami Fard, Alireza Heidari, Miadreza Shafie-khah, and João P.S. Catalão. Hybrid model using three-stage
algorithm for simultaneous load and price forecasting. Electric Power Systems Research, 165:214–228, 2018. ISSN 0378-7796. doi:
10.1016/j.epsr.2018.09.004.
Zhang Yang, Li Ce, and Li Lian. Electricity price forecasting by a hybrid model, combining wavelet transform, ARMA and kernel-based extreme
learning machine methods. Applied Energy, 190:291–305, 2017. ISSN 0306-2619. doi: 10.1016/j.apenergy.2016.12.130.
Javad Olamaee, Mohsen Mohammadi, Alireza Noruzi, and Seyed Mohammad Hassan Hosseini. Day-ahead price forecasting based on hybrid
prediction model. Complexity, 21(S2):156–164, 2016. doi: 10.1002/cplx.21792.
Jinliang Zhang, Zhongfu Tan, and Yiming Wei. An adaptive hybrid model for short term electricity price forecasting. Applied Energy, 258:114087,
2020. ISSN 0306-2619. doi: 10.1016/j.apenergy.2019.114087.
Rodrigo A. de Marcos, Antonio Bello, and Javier Reneses. Electricity price forecasting in the short term hybridising fundamental and econometric
modelling. Electric Power Systems Research, 167:240–251, 2019. ISSN 0378-7796. doi: 10.1016/j.epsr.2018.10.034.
Virginia Gonzalez, Javier Contreras, and Derek W. Bunn. Forecasting Power Prices Using a Hybrid Fundamental-Econometric Model. IEEE
Transactions on Power Systems, 27(1):363–372, 2012. doi: 10.1109/TPWRS.2011.2167689.

Watermeyer et al. Page 32 of 38

 A hybrid model for day ahead electricity price forecasting

Thomas Möbius, Mira Watermeyer, Oliver Grothe, and Felix Muesgens. Enhancing Energy System Models Using Better Load Forecasts. arXiv,
2023. doi: 10.48550/ARXIV.2302.11017.
Katarzyna Maciejowska and Jakub Nowotarski. A hybrid model for GEFCom2014 probabilistic electricity price forecasting. International Journal
of Forecasting, 32(3):1051–1056, 2016. ISSN 0169-2070. doi: 10.1016/j.ijforecast.2015.11.008.
Rafał Weron and Adam Misiorek. Forecasting spot electricity prices: A comparison of parametric and semiparametric time series models.
International Journal of Forecasting, 24(4):744–763, 2008. ISSN 0169-2070. doi: 10.1016/j.ijforecast.2008.08.004. Energy Forecasting.
Xia Chen, Zhao Yang Dong, Ke Meng, Yan Xu, Kit Po Wong, and H. W. Ngan. Electricity Price Forecasting With Extreme Learning Machine and
Bootstrapping. IEEE Transactions on Power Systems, 27(4):2055–2062, 2012. doi: 10.1109/TPWRS.2012.2190627.
Can Wan, Zhao Xu, Yelei Wang, Zhao Yang Dong, and Kit Po Wong. A Hybrid Approach for Probabilistic Forecasting of Electricity Price. IEEE
Transactions on Smart Grid, 5(1):463–470, 2014. doi: 10.1109/TSG.2013.2274465.
Mehdi Rafiei, Taher Niknam, and Mohammad Khooban. Probabilistic electricity price forecasting by improved clonal selection algorithm and
wavelet preprocessing. Neural Computing and Applications, 28, 12 2017. doi: 10.1007/s00521-016-2279-7.
Abbas Khosravi, Saeid Nahavandi, and Doug Creighton. Quantifying uncertainties of neural network-based electricity price forecasts. Applied
Energy, 112:120–129, 2013. ISSN 0306-2619. doi: 10.1016/j.apenergy.2013.05.075.
Jun Hua Zhao, Zhao Yang Dong, Zhao Xu, and Kit Po Wong. A Statistical Approach for Interval Forecasting of the Electricity Price. IEEE
Transactions on Power Systems, 23(2):267–276, 2008. doi: 10.1109/TPWRS.2008.919309.
Ming Zhou, Z. Yan, Y.X. Ni, Gengyin Li, and Y. Nie. Electricity price forecasting with confidence-interval estimation through an extended arima
approach. Generation, Transmission and Distribution, IEE Proceedings-, pages 187 – 195, 04 2006. doi: 10.1049/ip-gtd:20045131.
Anastasios Panagiotelis and Michael Smith. Bayesian density forecasting of intraday electricity prices using multivariate skew t distributions.
International Journal of Forecasting, 24:710–727, 10 2008. doi: 10.1016/j.ijforecast.2008.08.009.
Hans Manner, Farzad Alavi Fard, Armin Pourkhanali, and Laleh Tafakori. Forecasting the joint distribution of Australian electricity prices using
dynamic vine copulae. Energy Economics, 78:143–164, 2019. ISSN 0140-9883. doi: https://doi.org/10.1016/j.eneco.2018.10.034.
Oliver Grothe, Fabian Kächele, and Fabian Krüger. From point forecasts to multivariate probabilistic forecasts: The Schaake shuffle for day-ahead
electricity price forecasting. Energy Economics, page 106602, 2023. ISSN 0140-9883. doi: 10.1016/j.eneco.2023.106602.
Jakub Nowotarski and Rafał Weron. Computing electricity spot price prediction intervals using quantile regression and forecast averaging.
Computational Statistics, 30:791–803, 9 2015. ISSN 16139658. doi: 10.1007/S00180-014-0523-0/FIGURES/3.
Adam Misiorek, Stefan Trueck, and Rafal Weron. Point and Interval Forecasting of Spot Electricity Prices: Linear vs. Non-Linear Time Series
Models. Studies in Nonlinear Dynamics & Econometrics, 10(3), 2006. doi: 10.2202/1558-3708.1362.
Grzegorz Dudek. Multilayer perceptron for GEFCom2014 probabilistic electricity price forecasting. International Journal of Forecasting, 32(3):
1057–1060, 2016. ISSN 0169-2070. doi: 10.1016/j.ijforecast.2015.11.009.
Katarzyna Maciejowska, Jakub Nowotarski, and Rafał Weron. Probabilistic forecasting of electricity spot prices using factor quantile regression
averaging. International Journal of Forecasting, 32(3):957–965, 2016. ISSN 0169-2070. doi: 10.1016/j.ijforecast.2014.12.004.
Rafał Weron. Modeling and forecasting electricity loads and prices: a statistical approach. Wiley finance series. Wiley & Sons, Chichester [u.a.],
2006. ISBN 047005753X; 9780470057537.
Jakub Nowotarski and Rafał Weron. Merging quantile regression with forecast averaging to obtain more accurate interval forecasts of Nord Pool
spot prices. In 11th International Conference on the European Energy Market (EEM14), pages 1–5, 2014. doi: 10.1109/EEM.2014.6861285.
Grzegorz Marcjasz, Bartosz Uniejewski, and Rafał Weron. Probabilistic electricity price forecasting with NARX networks: Combine point or
probabilistic forecasts? International Journal of Forecasting, 36(2):466–479, 2020. ISSN 0169-2070. doi: 10.1016/j.ijforecast.2019.07.002.
Bartosz Uniejewski, Grzegorz Marcjasz, and Rafał Weron. On the importance of the long-term seasonal component in day-ahead electricity
price forecasting: Part II — Probabilistic forecasting. Energy Economics, 79:171–182, 2019. ISSN 0140-9883. doi: 10.1016/j.eneco.2018.02.007.
Energy Markets Dynamics in a Changing Environment.
Bartosz Uniejewski and Rafał Weron. Regularized quantile regression averaging for probabilistic electricity price forecasting. Energy Economics,
95:105121, 2021. doi: 10.1016/j.eneco.2021.105121.
Sandbag. Co2 emission allowance, 2020. URL https://sandbag.be/index.php/carbon-price-viewer/. Accessed on 03-01-2022.
Regelleistung.net. List of tenders capacity, 2018. URL https://www.regelleistung.net/ext/. Accessed on 12-12-2021.
Andreas Schröder, Friedrich Kunz, Jan Meiss, Roman Mendelevitch, and Christian von Hirschhausen. Current and Prospective Costs of Electricity
Generation until 2050. DIW Data Documentation, 68, 2013. ISSN 1861-1532.
Open Power System Data. Data Package Weather Data. Version 2020-09-16., 2020a. URL https://doi.org/10.25832/weather_data/
2020-09-16. Accessed on 15-05-2020.
ENTSO-E Transparency Platform. Total Load - Day Ahead / Actual, 2021a. URL https://transparency.entsoe.eu/. Accessed on 20-12-
2021.
Destatis Statistisches Bundesamt. Erzeugerpreise gewerblicher Produkte (Inlandsabsatz). Preise für leichtes Heizöl, Motorenbenzin und
Diesel, 2021. URL https://www.destatis.de/DE/Themen/Wirtschaft/Preise/Erzeugerpreisindex-gewerbliche-Produkte/
_inhalt.html. Accessed on 25-01-2021.
ENTSO-E. TYNDP 2018 Scenario Report, 2018. URL https://tyndp.entsoe.eu/tyndp2018/scenario-report. Accessed on 23-02-2023.
EEX. European Energy Exchange: Historic gas price data, 2021. Accessed on 03-01-2022.
BNetzA. Kraftwerksliste der Bundesnetzagentur, 2021. URL https://www.bundesnetzagentur.de/. Accessed on 03-01-2022.
UBA. Umweltbundesamt: Datenbank “Kraftwerke in Deutschland", 2020. URL https://www.umweltbundesamt.de/dokument/
datenbank-kraftwerke-in-deutschland. Accessed on 03-01-2022.
EBC. Europe Beyond Coal: European Coal Plant Database, 25 Jan 2021, 2021. URL https://beyond-coal.eu/database/. Accessed on
25-01-2021.
ENTSO-E Transparency Platform. Installed Capacities per Production Type, 2021b. URL https://transparency.entsoe.eu/. Accessed on
20-12-2021.

Watermeyer et al. Page 33 of 38

 A hybrid model for day ahead electricity price forecasting

Open Power System Data. Data Package National Generation Capacity. Version 2019-12-02., 2020b. URL https://doi.org/10.25832/
national_generation_capacity/2019-12-02. Accessed on 15-05-2020.
European Commission. Eurostat Statistics Database, 2021. URL https://ec.europa.eu/eurostat/data/database. Accessed on 12-12-
2021.
ENTSO-E Transparency Platform. Actual Generation per Production Type, 2021c. URL https://transparency.entsoe.eu/. Accessed on
20-12-2021.
ENTSO-E Transparency Platform. Forecasted Transfer Capacities - Day Ahead, 2021d. URL https://transparency.entsoe.eu/. Accessed
on 20-12-2021.
JAO Joint Allocation Office. ATC for Shadow Auction, 2021. URL https://www.jao.eu/implict-allocation. Accessed on 20-12-2021.
ENTSO-E Transparency Platform. Unavailability of Production and Generation Units, 2021e. URL https://transparency.entsoe.eu/.
Accessed on 20-12-2021.
ENTSO-E Transparency Platform. Generation Forecast - Day ahead, 2021f. URL https://transparency.entsoe.eu/. Accessed on 20-12-
2021.
ENTSO-E Transparency Platform. Day-ahead prices, 2021g. URL https://transparency.entsoe.eu/. Accessed on 20-12-2021.
Mark Hellwig. Entwicklung und Anwendung parametrisierter Standard-Lastprofile. Dissertation, Technische Universität München, 2003.
Dissertation, Technische Universität München.
DENA. dena-Netzstudie II. Integration erneuerbarer Energien in die deutsche Stromversorgung im Zeitraum 2015 – 2020 mit Ausblick
2025, 2010. URL https://www.dena.de/fileadmin/user_upload/Download/Dokumente/Studien___Umfragen/Endbericht_
dena-Netzstudie_II.PDF.
Friedrich Kunz, Mario Kendziorski, Wolf-Peter Schill, Jens Weibezahn, Jan Zepter, Christian R. von Hirschhausen, Philipp Hauser, Matthias Zech,
Dominik Möst, Sina Heidari, Björn Felten, and Christoph Weber. Electricity, heat, and gas sector data for modeling the german system. DIW
Data Documentation 92, 2017. URL http://hdl.handle.net/10419/173388.
Katarzyna Maciejowska, Weronika Nitka, and Tomasz Weron. Enhancing load, wind and solar generation for day-ahead forecasting of electricity
prices. Energy Economics, 99:105273, 7 2021. ISSN 01409883. doi: 10.1016/j.eneco.2021.105273.
Helmut Lütkepohl. New introduction to multiple time series analysis. Springer, Berlin, 2005. ISBN 3540401725; 3540262393; 9783540262398;
9783540401728.
Rob J. Hyndman and George Athanasopoulos. Forecasting: principles and practice. Otexts: Melbourne, Australia, Lexington, Ky., 2021. ISBN
0987507133. Accessed on 04-02-2022.
George E. P. Box, Gwilym M. Jenkins, Gregory C. Reinsel, and Greta M. Ljung. Time series analysis: forecasting and control. Wiley series
in probability and statistics. John Wiley and Sons Inc., Hoboken, New Jersey, fifth edition edition, 2015. ISBN 9781118675021. doi:
10.1111/jtsa.12194.
Roger Koenker and Gilbert Bassett. Regression Quantiles. Econometrica, 46(1):33, jan 1978. doi: 10.2307/1913643.
Oliver Grothe. A higher order correlation unscented Kalman filter. Applied mathematics and computation, 219(17):9033–9042, 2013. ISSN
0096-3003, 1873-5649. doi: 10.1016/j.amc.2013.03.019.
Felix Muesgens. Quantifying Market Power in the German Wholesale Electricity Market Using a Dynamic Multi-Regional Dispatch Model. The
Journal of Industrial Economics, 54(4):471–498, 2006b. doi: 10.1111/j.1467-6451.2006.00297.x.
Felix Muesgens and Karsten Neuhoff. Modelling Dynamic Constraints in Electricity Markets and the Costs of Uncertain Wind Output, 2006.
Grzegorz Marcjasz, Tomasz Serafin, and Rafał Weron. Selection of Calibration Windows for Day-Ahead Electricity Price Forecasting. Energies,
11:2364, 9 2018. ISSN 1996-1073. doi: 10.3390/EN11092364.
Victor DeMiguel, Lorenzo Garlappi, and Raman Uppal. Optimal versus naive diversification: How inefficient is the 1/n portfolio strategy. Review
of Financial Studies, 22(5):1915–1953, 2009. doi: 10.1093/rfs/hhm075.
Ludwig Kuntz and Felix Muesgens. Modelling start-up costs of multiple technologies in electricity markets. Mathematical Methods of Operations
Research, 66(1):21–32, 2007. ISSN 14322994. doi: 10.1007/s00186-007-0148-y.
Tilmann Gneiting, Fadoua Balabdaoui, and Adrian E Raftery. Probabilistic forecasts, calibration and sharpness. Journal of the Royal Statistical
Society: Series B (Statistical Methodology), 69(2):243–268, 2007. doi: 10.1111/j.1467-9868.2007.00587.x.

A. Additional Analysis Material

This appendix provides additional figures for Chapters 5 and 6.

Watermeyer et al. Page 34 of 38

 A hybrid model for day ahead electricity price forecasting

Figure A.1: RMSE of the day-ahead price forecast generated with the LEAR benchmark model for base, peak and
off-peak hours.

Figure A.2: RMSE of the day-ahead price forecast generated with the LEAR benchmark model for hours at different
day-ahead price quantiles.

Watermeyer et al. Page 35 of 38

 A hybrid model for day ahead electricity price forecasting

Figure A.3: RMSE for price estimators for each hour of the day and each year after the energy system optimisation
step in [A
C/MWh].

Figure A.4: Mean price forecast errors of the individual sub-models and the combined sub-models for the hours of a
day in each year in [A
C/MWh].

Watermeyer et al. Page 36 of 38

 A hybrid model for day ahead electricity price forecasting

Figure A.5: Mean price forecast errors of the individual sub-models and the combined sub-models for the hours of a
week (including holidays, hour 168 to 192) in [A
C/MWh].

Watermeyer et al. Page 37 of 38

 A hybrid model for day ahead electricity price forecasting

B. Nomenclature Energy System Model

Watermeyer et al. Page 38 of 38

 A hybrid model for day ahead electricity price forecasting

Watermeyer et al. Page 39 of 38