This article has been accepted for publication in IEEE Transactions on Smart Grid.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSG.2025.3573166

> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <

Enhanced GAN Based Joint Wind-Solar-Load

Scenario Generation with Extreme Weather Labelling
Hongzhen Wang, Student Member, IEEE, Boyu Qin, Member, IEEE, Shidong Hong, Xi Xu, Yiwei Su, Tingxiang Lu
and Tao Ding, Senior Member, IEEE

Abstract—The high penetration of renewable energy and for power system operation risk estimation. The accuracy of
frequent extreme weather events bring significant uncertainty to short-term scheduling schemes and long-term planning
the operation and planning of power systems. The extreme decisions depends on the extreme scenario samples [3]. Due to
scenarios exhibit high impact and low probability characteristics, unclear classification of extreme scenarios and insufficient
posing a challenge in providing sufficient and reliable samples for samples, it is challenging to systematically consider extreme
power system decision-making. The existing scenario generation scenarios in power system scheduling and planning [4].
methods struggle to balance the generation efficiency and Extreme weather is the principal cause of the occurrence of
interpretability. The causality between meteorological factors extreme scenarios in the power systems [5]. Extreme weather
and power system operation has not provided effective guidance events such as typhoons, blizzards, cold waves, and high
for the extreme scenario generation. In this paper, firstly, the
temperatures impact the physical structure and operation rules
mechanism analysis and sample construction of extreme
of the power systems from multiple aspects including source,
scenarios were achieved through meteorological factor extraction,
grid and load [6]. The influence of extreme weather on power
meteorological-scenario causality test and scenario clustering.
Secondly, an enhanced GAN model considering training stability system is the coupling effect of various meteorological factors.
is established. The physical constraints of power systems are used For example, in a blizzard, higher wind speeds will force wind
to ensure the effectiveness of generated scenarios. Finally, a turbines to shut down, snow will cover photovoltaic panels
statistical analysis considering the individuals and the whole and make them unable to generate electricity, and persistent
samples is conducted. The accuracy evaluation method of the low temperatures will also cause power loads to rise [7]. To
generated scenarios is proposed. The robustness of the model is accurately depict the impact of extreme weather on power
validated based on accuracy. The results indicate that the system, extreme weather is generally quantified by various
generated scenario have statistical consistency and high accuracy. meteorological factors with different intensities [8]. These
meteorological factors include temperature, humidity,
Index Terms—Scenario generation, generative adversarial precipitation, irradiance, wind speed, etc. Different
networks, meteorological factors, accuracy evaluation. meteorological factors have different effects on various
components in power systems [9]. According to experience,
I. INTRODUCTION certain meteorological factors are strongly associated with
specific scenarios, such as wind speed to wind power,

H
igh penetration of renewable energy and frequent
extreme weather events render the operation of the irradiance to solar power, temperature to load power. In [10],
power system confronted with significant uncertainty, the meteorological sensitivity of daily wind and solar energy
presenting substantial challenges to scheduling and yields, and energy demand in a highly-renewable European
planning [1]. Extreme events exert a considerable impact and energy system is investigated. The relationship between
possess strong destructiveness, yet their occurrence probability extremely high residual load and weather in Japan is studied in
is low, giving rise to a limited number of scenario samples [2]. [11]. In [12], the effect of the humidity and temperature on
The issue of small samples in extreme scenarios directly leads power load is considered using a hierarchical clustering
to an insufficient basis for decision-making. Therefore, it is of method. Besides these primary meteorological factors, there
great significance to address the small sample problem and exist some secondary meteorological factors that also should
generate extreme scenarios of power system with high quality. be considered. The differential contribution of various
Extreme scenarios are mainly used for power system meteorological factors to the scenarios can refine the effect of
scheduling and planning. Extreme scenarios provide the basis extreme weather on the power system and enhance the
diversity of the power system scenarios.
T The influence of extreme weather on power system
his work was supported by the Headquarter Management Science and
Technology Project of State Grid Corporation of China under Grant 5700- scenarios guides the identification and classification of data as
202440341A-2-1-ZX. (Corresponding author: Boyu Qin.) a pre-work for scenario generation. The existing scenario
H. Wang, B. Qin, S. Hong, Y. Su and T. Ding are with the School of generation methods are sampling-based method, optimization-
Electrical Engineering, Xi’an Jiaotong University, Xi’an 710049, China (e- based method and forecasting-based method [13]. The
mail: [email protected]; [email protected]; [email protected];
[email protected]; [email protected]). sampling-based method requires the probability density
X. Xu and T. Lu are with the Power Grid Dispatching & Control function (PDF) of the possible distribution, and fit the
Technology Branch Company, State Grid Electric Power Research Institute distribution parameters based on the existing samples [14].
(NARI Group Corporation), Nanjing 211106, China (e-mail: Monte Carlo sampling (MCS) is a simple and efficient
[email protected]; [email protected]).
Color versions of one or more of the figures in this article are available method, but the PDF of random variables (temperature, wind
online at http://ieeexplore.ieee.org speed and irradiance, etc.) must be known [15]. Sample

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 27,2025 at 16:06:06 UTC from IEEE Xplore. Restrictions apply.
© 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted,
but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Smart Grid. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSG.2025.3573166

> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <

average approximation (SAA) is an improvement of MCS the true distribution of data by constructing generative models
method, which replaces the actual distribution with the and generating new samples that conform to the data
empirical distribution of the sample [16]. Markov-chain distribution [33]. The two most used and effective methods are
Monte Carlo (MCMC) method uses Markov chains to variational autoencoders (VAEs) and generative adversarial
characterize the transition characteristics of the scenarios [17]. networks (GANs). The VAE constructs a probabilistic model
Latin hypercube sampling (LHS) pertains to stratified through an encoder and a decoder, allowing the encoder to
sampling, which ensures that all sampling areas are evenly learn low-dimensional representations of data and the decoder
covered by sampling points [18]. The number of sampling to generate more realistic samples [34]. The GAN is trained
points does not need to be overly large to reflect the through a generator and a discriminator, allowing the
probability distribution of random variables and can avoid generator to generate realistic samples and the discriminator to
repetitive sampling[19]. LHS is mostly used for the static accurately determine whether the samples come from the real
scenario generation, while the correlation of each period is not data distribution [35].
adequately considered in the dynamic scenario generation Compared to VAE, which requires more prior knowledge or
[20]. Copula function is mainly used for the joint scenario labelled data to guide the training process, GAN learns
involving wind power, solar power, and load power. The complex sample distributions through adversarial training and
Copula function sampling method has no restriction on edge unsupervised learning methods, and has been widely studied
distribution, and can capture the nonlinear, asymmetrical and in power system scenario generation [36]. The current
tail correlation between variables [21]. However, the above research work on GAN mainly focuses on the above problems,
sampling-based methods require prior assumptions and and improves the model structure and model training process
complex probability distribution modelling, which may lead to [37]. In [38], a controllable GAN model with interpretability is
large parameter fitting errors, and render it challenging to proposed for renewable scenario generation. In [39], an
handle the complex distribution features of actual renewable improved Wasserstein GAN based wind power scenario
energy and load. generation method is proposed for multiple wind farms. In
A set of random scenarios with equal probability is obtained [40], a conditional style-based GAN is established for day-
by sampling-based method. A large number of these scenarios ahead scenario generation of renewable power. In [41],
are similar and do not provide differentiated information for federated learning and least square GAN are integrated for
power system operation and planning, but significantly reduce scenario generation in a privacy-preserving manner.
computational efficiency [22]. The optimization-based method The above improved GANs can effectively capture the
is an extension of sampling-based scenario generation, which temporal characteristics of renewable energy and power load,
cuts scenarios to improve the efficiency of scenario but it fails to fully reflect the causal relationship between
description [23]. The moment matching method uses statistical extreme weather and power system scenarios, and the
features such as expectation, standard deviation, skewness and interpretability of generated scenarios is insufficient. In
kurtosis to describe the features of random variables [24]. Addition, there are physical constraints between the
Through matrix transformation and cubic transformation of generation scenario and the actual power system operation.
random historical variables, the random features of historical The generated time series must strictly meet the technical
data are preserved in the generated scenarios [25]. Distance requirements of the power system for power boundaries, ramp
matching methods mainly include forward selection (FS), power, etc. The rationality of the generation scenario needs to
backward reduction (BR), clustering, etc. [26]. The FS method be further guaranteed.
selects typical scenarios one by one from massive scenarios. To fill the research gap mentioned above, a causal
The BR method eliminates unrepresentative scenarios one by assessment method for extreme weather and power system
one from massive scenarios [27]. The clustering method scenarios has been proposed. On this basis, an extreme
gathers massive scenarios into several categories and uses the scenario generation model based on the enhanced GAN is
representative samples of each category to compose typical constructed. The main contributions of this paper are as
scenarios [28]. The optimization-based method is essentially follow.
reducing scenarios and may ignore extreme scenarios with low 1) The effects of meteorological factors on extreme
probability but high impact. scenarios are quantified based on Granger causality test. The
The forecasting-based method uses historical data to train power time series were clustered using K-MDTSC and formed
the scenario generation model. This method pays more a scenario set with classification labels. The impact analysis of
attention to the temporal relationship of variables without extreme weather and meteorological factors on the power
considering their statistical features, avoiding complex system scenarios enhances the interpretability of the clustering
probability distribution modelling [29]. Autoregressive results and guide the selection and construction of extreme
moving average (ARMA) model is a classical forecasting- scenario samples set.
based method [30]. Autoregressive integrated moving average 2) An extreme scenario generation model with gradient
(ARIMA) model and Wavelet ARIMA are improved penalty and variable learning rate based on Wasserstein GAN
algorithms based on ARMA to deal with the time series with is established. The learning ability and convergence of the
non-stationary and non-Gaussian patterns [31]. However, for model are balanced. The power constraints are introduced into
complex time series, the accuracy of ARMA method is still the generation model. The generated model is trained under
affected by inaccurate transformation [32]. Machine learning the guidance of the classification labels to realize the
methods are a new class of non-parametric methods that learn directional generation of extreme scenarios.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 27,2025 at 16:06:06 UTC from IEEE Xplore. Restrictions apply.
© 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted,
but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Smart Grid. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSG.2025.3573166

> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <

3) A statistical feature evaluation method considering the where pmk (i )

is a proportion of class-k observations at node m.
individual and the whole of the generated scenarios is For the decision tree i, the importance of the meteorological
proposed. An accuracy evaluations method for the generated factor j in the node m can be expressed by the Gini Gain as
scenarios, encompassing accuracy of points and accuracy of follows.
scenarios, has been proposed. Considering the influence of
VIM (jmi ) = GI m(i ) − GIl(i ) − GI r(i ) (2)
noise and outliers, the robustness and adaptability of the
generation model are evaluated base on the accuracy where l and r denote the two daughter nodes of m.
evaluations method. If the meteorological factor j appears in the node set M in
The rest of this paper is organized as follows. In Section 2, the decision tree i, the Gini Gain of the meteorological factor j
the relationship between extreme weather and power system on the decision tree i is as follows.
scenarios is constructed, and samples are prepared for scenario VIM (ji ) =  VIM (jmi ) (3)
generation models. In Section 3, an extreme scenario mM

generation model based on GAN is established. In Section 4, Considering n decision trees in random forest, the
the proposed extreme scenario generation model is trained. In normalized contribution of meteorological factor j is the
Section 5, the statistical analysis and accuracy evaluation of importance score.

n
the generated scenario are conducted. The robustness of the VIM (j i )
model is also analyzed. Finally, Section 5 summarizes the VIM j = c i =1
(4)
 j =1  i =1VIM (ji )
n
main contents and key results of the work.
It is notable that feature importance merely indicates the
II. SAMPLE PREPARATION contribution of meteorological factors to extreme weather and
In this section, extreme weather is characterized by cannot reflect causality. The effects of meteorological factors
meteorological factors of different intensities. Based on on extreme scenarios are analyzed in the next subsection.
meteorological factors, the relationship between extreme B. Effects of Meteorology on Extreme Scenarios
weather and scenarios is established. The causality test of
meteorological factors and scenarios provides theoretic Meteorological factors are related to power system
foundation for scenario classification and selection. The operation scenarios, especially wind power, solar power and
extreme scenarios with labels integrate feature identification load. The combination of meteorological factors with different
information, which provide high-quality samples for scenario intensities reflects different types of extreme weather, which
generation. has different effects on the operation and planning of power
system. Meteorological factors such as temperature,
A. Extraction of meteorological factors for extreme weather precipitation, illumination, wind speed and snowfall are
The meteorological factors of extreme weather are diverse considered, and their effects on wind power, solar power, and
and complex in composition. In order to balance quantification loads will be quantified by correlation analysis.
accuracy and computational efficiency, it is necessary to Granger causality test is widely used to test the causality
extract meteorological factors that have a significant impact between two time series. For the meteorological time series X
on extreme weather. In addition, the quantification of and power time series Y, augment the historical value of one
meteorological factors also provides better interpretability for series to ascertain if it can enhance the prediction of the other
the impact of extreme weather on power system scenarios. series for determining whether there is a Granger causality.
Random forest is an ensemble learning method consisting The regression is estimated through the least squares method
of multiple decision trees. a portion of samples from the as follows.
k k
sample set is selected using bootstrap resampling method as a yt =  0 +   i yt −i +  i xt −i + u1,t (5)
training set. a decision tree is generated based on the sample i =1 i =1
set obtained by sampling. At each generated node, data The test uses the null hypothesis H0: the meteorological
features are selected randomly and non-repeatedly. According time series will not lead to the Granger distribution of the
to these features, the sample set is divided to find the optimal power time series. The null hypothesis is presented
division. The essence of feature extraction with random forest mathematically as follows.
is to evaluate the contribution or importance of each feature on H 0 : 1 =  2 = =  k = 0 (6)
each tree in the random forest. The Gini Impurity is an The Granger causality test is conducted through a
indicator measure the purity of dataset. Essentially, this constrained F test. Suppose the null hypothesis H0 holds. That
method measures how much the impurity within a node of a is, assuming that the parameters before each lag of X in the
decision tree decreases when a specific feature is used to split formula are all zero, perform separate regressions with and
the data. without each lag of X included. Record the sum of squares of
For a certain type of extreme weather, considering the residuals for as RSSU and RSSR, respectively. Then calculate
meteorological factors x1, x2, ... xc, the Gini Impurity at the the F statistic as follows.
node m of the decision tree i is shown as follows. ( RSS R − RSSU ) m
F= (7)
RSSU ( n − k )
K
GI m(i ) = 1 −  ( pmk )
(i ) 2
(1)
k =1 where m is the number of lag terms of X, n is the sample size,
and k is the number of parameters to be estimated in the

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 27,2025 at 16:06:06 UTC from IEEE Xplore. Restrictions apply.
© 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted,
but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Smart Grid. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSG.2025.3573166

> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <

unconstrained regression model including possible constant SSE =   d ( X N , CN ) (9)

terms and other variables. If the calculated F value is greater CCˆ X N C

than the critical value Fα(m, n-k) of the F distribution under a where C is a cluster and Ĉ is the set of all clusters.
given significance level α, the null hypothesis H0 can be
The causality test of meteorological factors and scenarios
rejected and it can be concluded that the meteorological time
enhances the interpretability of clustering. The clusters with
series is the Granger cause of the power time series.
extreme scenario characteristics are selected. The filtered
It is worth noting that the Granger causality test is based on
clustering results are labelled c for power system scenarios
stationary time series. The mean, variance and covariance of
such as wind power Xw, solar power Xs and load power Xl.
the stationary series do not change with time, and the
They are constructed as learning samples {Xw, Xs, Xl |c} for the
prediction is reasonable. The regression model based on non-
following scenario generation.
stationary sequence has the risk of spurious regression.
Therefore, before the Granger causality test, the unit root test
III. EXTREME SCENARIO GENERATION METHOD
should be performed on the stability of each time series.
Augmented Dickey-Fuller test (ADF test) is used in the unit The principle of extreme scenario generation is to learn the
root test. If several non-stationary time series become feature information of real samples and generate synthetic
stationary after a difference, and long-term equilibrium is samples similar to real samples. In order to achieve the small
confirmed by a cointegration test, the Granger causality test sample expansion and data augmentation of extreme scenarios, an
can be continued. Otherwise, there may be spurious regression improved GAN is used for the extreme scenario generation. The
between the time series. The flow chart of Granger causality structure, loss function and constraints of the network model are
test is shown in Fig. 1. proposed in this section.
Start
A. Conditional Generative Adversarial Networks
Input meteorological and The GAN includes two neural networks: generator G and
power time series No discriminator D. G and D compete until the generator has the
ability to generate synthetic samples similar to real samples.
Unit root test Difference the time series
However, there are two deficiencies in GANs. First, the
generated samples may focus on some modes, resulting in
mode collapse. Secondly, during the training process, it is easy
Time series smooth? Cointegration test to occur that one of the generators or discriminators is rapidly
Yes enhanced, and it is difficult to achieve Nash equilibrium.
Yes
Granger causality test Long-term equilibrium? Based on the GANs, the conditional GANs (CGANs) add
No labels to the input data of the generator and discriminator.
End Spurious regressions Data labels are used as additional information to guide the
Fig. 1. Flow chart of Granger causality test. generation of samples. This combination of supervised
learning and semi-supervised learning ensures the diversity of
C. Scenario Classification and Selection generated sample and can also specify the required sample
Meteorological factors have higher dimensions than power types.
system scenarios, and their characteristics are more The structure of CGAN is shown in Fig. 2. Suppose that the
significant. The classification of scenarios based on real sample is x obeying X~X distribution, and c is the label
meteorological factors can further clarify the characteristics of
of each sample. A group of input latent variables z under a
scenarios and guide the generation of extreme scenarios.
K-Multi-Dimensional Time Series Clustering Algorithm (K- known distribution Z~Z is fed into G. The synthetic samples
MDTSC) is used to perform time series clustering on the G(Z) are output of G. D takes G(z) or a real sample x as input,
above multi-dimensional meteorological factors. Based on the and measures to how realistic the input sample by outputting a
framework of K-means clustering algorithm, K-MDTSC continuous scalar. The function of D is as follows.
updates the distance calculation for multi-dimensional time
series. Taking N as the dimension of clustering samples and T D ( x | c )

preal =  (10)
D (G ( z | c ))
as the time scale of clustering samples, the distance 
calculation of K-MDTSC is formulated as follows.
N T
Based on the defined D and G, their loss functions are
d ( X N , CN ) = L   X n (t ) − Cn (t )
L
(8) constructed. The generator loss function LG reflects the degree
n =0 t =0 to which the generated samples are identified as real in D. The
where L denotes the metric distance. XN and CN are the smaller the LG, the closer the distribution of generated samples
synchronous multi-dimensional meteorological time series. is to X. The discriminator pursues a stronger ability to
Like traditional clustering algorithms, the square of the sum of
the distance between the sample point and its cluster center, distinguish the difference between generated samples and real
that is, within-cluster sum of squared errors (SSE) or cluster samples. The smaller LD, the more significant the distribution
inertia, is used to evaluate the performance of the algorithm difference between generated samples and real samples. The
and determine the optimal number of clusters. The calculation loss functions of G and D are as follow.
formula of SSE is as follows.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 27,2025 at 16:06:06 UTC from IEEE Xplore. Restrictions apply.
© 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted,
but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Smart Grid. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSG.2025.3573166

> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <

LG = − Z Z
 D ( G ( Z | c ) )  (11) penalty increases as the distance from 1 increase. Combine
(16) and (18) to form the loss function LDP as follows.
LD = − X X
 D ( X | c )  + Z Z  D ( G ( Z | c ) )  (12) LDP = X X  D ( X | c ) 
(19)
− Z Z  D ( G ( Z | c ) )  +  L p
LG is opposite to the second term of LD. Combine (11) and
(12) to form the min-max game values function V(G, D). The
game value function is as follows. where λ is the weight of gradient penalty. The loss function of
min max V ( G, D ) = X X  D ( X | c )  deep learning is non convex and can only be adjusted by
 (G )  ( D ) experience. If λ is too small, the gradient penalty is not enough
(13)
− Z Z  D ( G ( Z | c ) )  to constrain the gradient of the discriminator, which may lead
to unstable training (such as gradient explosion). If λ is too
where θ(G) and θ(D) are the neural networks parameters of G
large, the gradient penalty term dominates the loss function
and D, respectively. G and D optimize their parameters in the
and weakens the performance of the discriminator, resulting in
game with each other, and finally achieve Nash equilibrium.
the decline of discrimination ability and the stagnation of
RMSProp and Adam are the essential adaptive optimizers for
generator learning. The model that introduces Wasserstein
gradient descent. The gradients of loss functions are as follow.
distance and gradient penalty based on CGAN is referred to as
g ( D ) =  ( D )  − D ( X | c ) + D ( G ( Z | c ) )  (14) WGAN-GP.
g ( G ) =  ( G )  D ( G ( Z | c ) )  (15) The D in GAN generally need to learn faster than G. The D
must learn the new patterns before they are transferred to G.
These gradients are combined with a given learning rate for Otherwise, the D will be driven to the invalid patterns by the
parameter optimization. G. In addition, the learning rate needs to balance the learning
effect and convergence. A separate variable learning rate is
Encoding Historical Gradient Discriminator
Labels Samples Penalty Loss used to improve the performance of the model. The
[0 1 0 0 0] Update constructed learning rates for the parameter optimization of G
[0 0 1 0 0]
[0 0 0 0 1] and D are as follows.
Output
i = 0i −r (20)
Fake Wasserstein

Generator Generated
Real Distance i = 0i −r (21)
Noise Input Samples
Discriminator where αi and βi are the parameters of G and D in the ith
Generator iteration, respectively. r is damping coefficient and r  (0,1] .
Constraints
Loss
α0 and β0 are the initial learning rate, respectively. β0 is larger
Update (Adaptive Optimizer)
than α0. The initial learning rate usually starts from 10-4,
Fig. 2. Extreme scenario generation framework. gradually increases until the loss fluctuates, and then falls
back to the stable value [42]. As the iteration progresses, the
B. Data Augmentation Application
learning rate gradually decreases and the convergence is
In order to solve the problem of small samples in extreme enhanced.
scenarios of power systems, a data augmentation method The generator's objective is to get the generated sample
based on GAN is proposed. Wasserstein distance is chosen as distribution as near to the real sample distribution as possible.
the loss function of D. Considering that the minimization of However, power system operation scenarios not only have
Wasserstein distance is difficult to achieve, its dual form is requirements for probability distribution, but also have strict
obtained by using Kantorovich-Rubinstein duality. restrictions on the physical implications. For example, the
W ( X , Xˆ ) = sup X X  D ( X | c )  installed capacity of the power generation system has
f D L 1 constraints on the maximum power, and the equipment
(16)
− Z Z  D ( G ( Z | c ) )  characteristics also have constraints on the rate of power
change. To ensure the practical operation requirements of the
where Xˆ = D(G(Z | c)) and Xˆ ~ Xˆ
. The discriminator generated scenarios, the constraints and corrections on the
function fD needs to meet 1-Lipschitz to remain gradients generated samples are added after the generator. The
smooth and avoid gradient explosion, that is constraints of power range and climbing rate are shown as
follow.
f D ( x1 ) − f D ( x2 )  x1 − x2 (17)
0  xˆt  xlim , t τ (22)
The 1-Lipschitz constraint of fD is imposed on the loss
xˆt +t − xˆt  x max
(23)
function in the form of a penalty gradient. After normalization
and expectation, the penalty term is as follows. where xlim is the upper limits of generated samples. Δt is the
time interval for the climbing rate calculation. Δxmax is the
Lp = X   x D ( x ) 2 − 1  ( )
2
(18) maximum power change in Δt.
X 
 
Comprehensive sample preparation as well as generation
where x = tx + (1 − t )G ( z ) and t ~ (0,1) . This sampling
model construction and training methods, the framework of
method ensures comprehensive coverage of the sample space extreme scenario generation is shown in Fig. 3.
and computational feasibility. Eq. (18) indicates that a penalty
occurs when the gradient of function D is not 1, and the

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 27,2025 at 16:06:06 UTC from IEEE Xplore. Restrictions apply.
© 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted,
but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Smart Grid. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSG.2025.3573166

> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <

load data for the same time and location is collected from
Meteorological disaster records Raw meteorological data
OCCTO website (www.occtonet3.occto.or.jp). The interval of
the original time series is 1 hour. Time resampling and
Corresponding Time series segments of extreme weather normalization organize data into a time series with intervals of
time series
segments of
Random Forest 5 minutes, ranging from 0 to 1.
extreme scenarios Extraction of key meteorological factors Based on the meteorological disaster records released by the
K-MDTSC Granger causality test local meteorological department, time series segments of
Causality test of meteorology and scenarios
various extreme weather events are initially sorted out. These
time series segments are grouped together. Random forest is
Selection
principles Normal scenarios used to analyze the importance of various meteorological
factors in these extreme weather events. The extreme weather
Scenario clustering Cluster selection Extreme scenarios
has a clear seasonal and regional distribution. For regions with
specific extreme weather events (e.g. blizzards in HS and HK,
Data
high temperature in KS, SG and HS), the corresponding
Classified
Augmentation Initializing
output of
meteorological factors contribute in a relatively orderly and
and stable manner. Temperature, precipitation, illumination, wind
Power system extreme
training
Constraints Build GAN model scenarios speed and snowfall are selected as the key meteorological
factors. The causality of meteorological factors on power
Fig. 3. Framework of extreme scenario generation. system scenarios can be reflected by the p-value of the
Granger causality test results in TABLE I. Wind speed and
IV. RESULTS OF SCENARIO GENERATION snowfall have a significant impact on wind power,
The proposed extreme scenario generation method is temperature, light, and wind speed have a significant impact
implemented in this section. First, an extreme scenario set on solar power, and temperature and light have a significant
under extreme weather events is constructed by impact on load power.
meteorological factors analysis. Second, historical scenario
samples are input into the generation model to achieve the TABLE I
classification generation of extreme scenarios. Finally, the RESULTS OF GRANGER CAUSALITY TEST
performance of the generation model is evaluated through Meteorological p-value
statistical metrics and robustness analysis. factors Wind power Solar power Load
Temperature 0.4599 0.0000 0.0000
A. Data Description and Preparation
Precipitation 0.6050 0.9779 0.2287
To focus on extreme scenario research, the raw data is taken
from three years of wind power, solar power, load power, and Illumination 0.8246 0.0296 0.0000
meteorological data in Japan from 2021 to 2023. The locations Wind speed 0.0052 0.1079 0.4306
for data collection include Kyushu (KS), Shikoku (SG), Snowfall 0.0041 0.9082 0.1564
Honshu (HS), and Hokkaido (HK). The meteorological and
renewable energy data for the corresponding time and location
is collected from Ninja website (www.renewables.ninja). The
Temperature Precipitation Snowfall Illumination Windspeed Wind power Solar power Load power
Power (p.u.) Power (p.u.) Power (p.u.) Power (p.u.) Power (p.u.)
Class 1
Class 2
Class 3
Class 4
Class 5

Time (h) Time (h) Time (h) Time (h) Time (h) Time (h) Time (h) Time (h)

Fig. 4. Clustering and screening results of extreme scenarios.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 27,2025 at 16:06:06 UTC from IEEE Xplore. Restrictions apply.
© 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted,
but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Smart Grid. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSG.2025.3573166

> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <

The optimal number of clusters for historical scenario

samples is determined to be 20 based on the elbow method.

D(x)/D(G(z))

D(x)/D(G(z))
Based on the analysis of the importance of meteorological
factors on extreme weather, 5 representative extreme scenarios
were selected from 20 historical scenarios. The sample sizes
of the 5 extreme scenarios are 41, 86, 37, 36 and 28. The
scenario classification is shown in Fig. 4. These extreme
scenarios can respectively represent the 5 extreme weather
(a) λ=0 (b) λ=15
conditions of typhoons, high temperatures, cold waves,
blizzards, and rainstorms. In addition, high temperature and

D(x)/D(G(z))

D(x)/D(G(z))
high load correspond to Class 1 and Class 2, low wind speed
and low wind power correspond to Class 2 and Class 5, and
low light and low solar power correspond to Class 3, Class 4,
and Class 5. The causal test results based on meteorological
factors and power scenarios can basically verify the
effectiveness of the classification results.
(c) λ=30 (d) λ=100
B. Model Architecture and Implementation Fig. 6. Training evolution for GANs with different weight of
The proposed model is based on a more stable and deeper deep gradient penalty.
convolutional generative adversarial network (DCGAN). In the
generator, a random noise vector z with label c is reshaped into a To verify the effectiveness of the proposed model, CGAN
tensor through the fully connected layer. The transposed and WGAN-GP are used as references for runtime evaluation.
convolution is used for tensor transformation, followed by batch CGAN, WGAN-GP and the proposed GAN demonstrate
normalization and ReLU activation. The transposed convolution progressive improvements in model convergence. The
is applied twice in the generator, and finally tanh activation is comparison of running time is shown in TABLE II. Compared
applied in the output layer. Convolution is used in the with the CGAN, WGAN-GP achieves significant convergence
discriminator for tensor transformation, followed by batch enhancement through the strategic incorporation of the
normalization and LeakyReLU activation. After two Wasserstein distance metric and gradient penalty mechanism.
convolutions, the full connection layer is used, the sigmoid The optimization of the gradient penalty weight and learning
activation function is input to calculate preal. In the process of rate of the proposed GAN further reduces epochs and
data transformation, the generator depth decreases, which is up- convergence time of the model.
sampling, while the discriminator depth increases, which is
down-sampling. The model architecture is shown in Fig. 5. TABLE II
The proposed GANs model is built on the Tensorflow platform COMPARISON OF THE RUNTIME
and programmed by python. All the results are implemented on CGAN WGAN-GP Proposed GAN
AMD Ryzen 5 3600 CPU with 16 GB RAM. Epochs 419 114 92
Convergence
158.64 38.34 29.11
time (s)
Decoder

100
(z+c)
G(z)
Reshape DeConv DeConv
D. Scenario Generation
Each class of scenarios is set to generate 200 samples, a total
Encoder

preal of 1000. The time series comparison between the generated

Conv Conv Sigmoid
scenarios and the real samples is shown in Fig. 7. There are
significant features between various classes of generated
DeConv2D BN ReLU Conv2D BN LeakyReLU scenarios, and they have good interpretability for extreme
Deconvolutional Layer Convolutional Layer weather. For example, the high temperature weather
Fig. 5. Model architecture for extreme scenario generation. represented by Class 2 has the highest load level, and the
rainstorm weather represented by Class 5 has the lowest wind
C. Training Results
power and solar power. In the same class of scenario, the
To verify the effectiveness of the gradient penalty and obtain generated scenario can basically track the patterns and trends
the optimal weight, the training evolution for GANs with of real samples, with good consistency.
different weight of gradient penalty is shown in Fig. 6. The loss
function of deep learning is non convex and can only be adjusted V. PERFORMANCE ASSESSMENT
by experience. There is good convergence when the weight of
gradient penalty is 15. The lack of gradient penalty can lead to A. Statistical Analysis
difficulties in model convergence. However, when the weight of The statistical features of the extreme scenario time series
gradient penalty is too large, it will limit the discriminative are analyzed from the individual and the whole of the
ability of the discriminator, causing deviation in the convergence generated scenarios. At the individual level, the generated
of the model. scenarios need to satisfy the temporal correlation of source

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 27,2025 at 16:06:06 UTC from IEEE Xplore. Restrictions apply.
© 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted,
but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Smart Grid. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSG.2025.3573166

> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <

and load. The autocorrelation coefficient (ACF) and partial

autocorrelation coefficient (PACF) can analyze and evaluate

PACF
the temporal correlation of the generated scenarios. The

ACF
calculation formula of the autocorrelation coefficient R(k) is as
follows.
E  ( xt −  )( xt + k −  )
R( k ) = (24) Lags (5 mins) Lags (5 mins)
E ( xt −  )2 (a) Wind power
where k is the lag number of time series, μ is the mean value
of the time series.

PACF
ACF
Lags (5 mins) Lags (5 mins)
(b) Solar power

PACF
ACF
Lags (5 mins) Lags (5 mins)
(c) Load power
Fig. 8. ACF (left) and PACF (right) of the real samples and
generated scenarios.

CDF
PDF

Power (p.u.) Power (p.u.)

(a) Wind power

Fig. 7. Time series comparison of real samples and generated

CDF
PDF

scenarios.

The PACF refers to the correlation of xt-n and xt under the

condition of n-1 random variables xt-1, …, xt-n+1 in the given Power (p.u.) Power (p.u.)
(b) Solar power
time series, that is, after eliminating the influence of n-1
random variables, the calculation formula is as follows.
E ( xt − Eˆ ( xt ))( xt − n − Eˆ ( xt − n )) 
CDF
PDF

P ( n) = (25)
E ( xt − Eˆ ( xt )) 2 E ( xt − n − Eˆ ( xt − n )) 2
where Eˆ ( xt ) = E ( xt | xt −1 , , xt − n +1 ) is the correlation between
xt-1, …, xt-n+1 and xt. Power (p.u.) Power (p.u.)
(c) Load power
The PACF measures the direct relationship between a
sequence and one of its lagged terms, eliminating the Fig. 9. PDF (left) and CDF (right) of the real samples and
interference of intermediate lagged terms. It is most useful for generated scenarios.
identifying the order of an autoregressive model. Sample
partial autocorrelations that are significantly different from 0 ACF and PACF of real samples and generated scenarios are
indicate lagged terms of x that are useful predictors of xt. It is shown in Fig. 8. The ACF of the real samples and the
important that the choice of the order makes sense. The time generated scenarios are very close, and the downward trend is
series power has strong longitudinal similarity, so the ACF of consistent. Compared with wind and load power, the ACF of
the time series gradually decreases with the increase of lags, solar power decreases rapidly with the increase of lag number.
while the PACF has strong correlation only when the interval This is partially caused by the greater volatility of solar power
is short. The validity of the generated scenarios can be verified output. The generated scenarios and the real samples have
by comparing the correlation coefficients. high consistency considering PACF. The generated and real

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 27,2025 at 16:06:06 UTC from IEEE Xplore. Restrictions apply.
© 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted,
but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Smart Grid. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSG.2025.3573166

> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <

PACFs of wind power tail off to zero. The PACFs of solar TABLE III
power cut off after the 1st lag, while the PACFs of load power MATCHING QUALITY OF REAL AND GENERATED SCENARIOS
cut off after the 4th lag. WGAN- Proposed
The cumulative distribution function (CDF) of discrete Indictor CGAN
GP GAN
random variables is used to analyze the different power Class 1 0.05068 0.03391 0.03468
distributions of the generated time series. CDF is an integral of Class 2 0.05582 0.03474 0.03359
PDF and is a non-decreasing function. This feature ensures Class 3 0.05154 0.03555 0.03223
that the probability of a random variable being less than or MAE
Class 4 0.05432 0.03666 0.03578
equal to a certain value always increases. In the generated Class 5 0.06235 0.03884 0.03430
scenarios, due to the introduction of physical constraints, the Average 0.05494 0.03594 0.03412
power can be guaranteed to be between 0 and the power limit. Class 1 0.06451 0.04405 0.04471
CDF of the generated time series is a monotonically rising Class 2 0.07395 0.04585 0.04305
curve from 0 to maximum power on the x-axis and from 0 to 1 Class 3 0.06530 0.04592 0.04214
on the y-axis. RMSE
Class 4 0.06796 0.04687 0.04496
The PDF and CDF of extreme scenarios is shown in Fig. 9. Class 5 0.07931 0.05048 0.04336
The wind power and solar power is mainly concentrated Average 0.07020 0.04663 0.04365
around 0, and this phenomenon is more pronounced in solar Class 1 11.4588 7.96249 7.61465
power. The load power is mainly concentrated between 0.1 Class 2 11.9185 8.11548 7.84534
and 0.7. The load power distribution is more uniform, MAPE Class 3 9.96405 6.91284 6.07786
resulting in a smooth upward trend of CDF. The probability (%) Class 4 10.2583 6.91883 6.50165
distribution of the generated scenarios and the real samples is Class 5 15.8600 9.92137 9.44831
basically consistent. Average 11.8919 7.96620 7.49756
To accurately evaluate the matching quality of the real
scenario and the generated scenario, the mean absolute error There are two accuracy metrics proposed. First, the accuracy
(MAE), root mean square error (RMSE) and mean absolute of points Ap. The validity of each point in the generated time
percentage error (MAPE) are introduced. MAE provides a series is considered. Ap is determined by the ratio of the
basic error scale measurement. RMSE can highlight the number of generated points within a valid range to the total
impact of large errors and help to find outliers. MAPE can number of generated points. In the generated scenarios, the
reveal the relative size of errors and is suitable for evaluating length of the time series for each wind, solar and load power is
the processing ability of different proportion errors. The the same, so the total number of generated points is fixed. The
combination of these three evaluation indexes can achieve a calculation for Ap is as follows.
balance between absolute error, large error sensitivity, and

3
relative error, and is suitable for the comprehensive j =1
N pv , j
Ap = 100% (26)
performance evaluation of extreme scenarios. The results are 3N s L
shown in TABLE III. The matching quality under WGAN-GP
where Npv,j is the number of generated points within a valid
is significantly higher than that under CGAN. The matching
range, j is used to distinguish wind, solar and load power. Ns is
quality of the proposed GAN in Class 3 and Class 5 is
the number of generated scenarios. L is the length of each time
significantly higher than that of WGAN-GP, and the other is
series of generated scenarios.
not much different. This is due to the introduction of physical
Second, the accuracy of scenarios As. A valid scenario
constraints in the proposed GAN. The proposed GAN has a
requires the time series of wind, solar and load power to be
significant advantage in overall matching quality compared to
within the power range. As reflects the ability of the model to
CGAN and WGAN-GP.
generate complete valid scenarios. As is determined by the
B. Accuracy of the Generated Scenarios ratio of the number of valid scenarios to the total number of
Visual inspection and statistical analysis can qualitatively generated scenarios. The calculation for As is as follows.
test the similarity and diversity between generated scenarios N
As = sv 100% (27)
and real samples. Statistical indicators are designed to reveal Ns
the characteristics of a single time series. They verify the where Nsv is the number of generated valid scenarios. It can be
ability of the generation model to reproduce the statistical concluded that As is a more stringent metric than Ap. Even if
characteristics of the power system scenarios. Accuracy is a some points of the time series are not within the power range,
quantitative indicator for evaluating generated scenarios. In or some types of power series is invalid, the complete scenario
the real sample shown in Fig. 4, the wind power, solar power, is considered invalid.
and load of each scenario form corresponding power ranges. The accuracy of generated scenarios is shown in Fig. 10.
These power ranges reflect the scenario characteristics under The proposed GAN has certain advantages in accuracy over
their corresponding extreme weather conditions. When the CGAN and WGAN-GP. WGAN-GP ensures that it satisfies
generated scenarios are not within or partially outside these the 1-Lipschitz condition by penalizing the gradient of the
power ranges, they are invalid. The accuracy index focuses on discriminator. In addition to this basic function, gradient
the effectiveness and applicability of the generated scenarios penalty makes the discriminator not only focus on
in the power system. distinguishing between real and generated samples, but also

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 27,2025 at 16:06:06 UTC from IEEE Xplore. Restrictions apply.
© 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted,
but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Smart Grid. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSG.2025.3573166

> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <

pay more attention to maintaining the smoothness of the in advance. Gaussian noise following normal distribution is
discriminant function. This constraint can make the added to the samples with a certain signal-to-noise ratio (SNR).
discriminator more universal to the input data, rather than The noise is completely covered on the real samples. The
overfitting to the local features of the training samples. For smaller the SNR, the greater the noise, and the worse the
generator, providing stable and accurate feedback can help quality of the generated scenarios. The maximum SNR is set
them learn higher quality samples and improve their diversity to 10 dB. Different proportions of outliers are also considered
and quality. The proposed GAN optimizes the training in the modified samples. There are many values of 0 or close
parameters and adds physical constraints. When invalid data to 0 in the solar power series. In the real sample, the points in
appears in the generated scenarios, the generator corrects it the time series are randomly replaced by 1. As is less
through range and ramp restrictions. Therefore, compared to distinguishable in robustness evaluation. The robustness of the
WGAN-GP, there is a significant improvement in accuracy of generation model is evaluated by Ap.
the overall generated samples. Ap has a more refined
evaluation of the scenarios than As, resulting in little difference TABLE IV
between various methods. As evaluated the complete scenarios GENERATION QUALITY OF NORMAL AND EXTREME SCENARIOS
and found that the proposed GAN has a more significant WGAN- Proposed
advantage in valid scenario generation. Indictor CGAN
GP GAN
Normal 0.05114 0.03682 0.03360
MAE
Extreme 0.05494 0.03594 0.03412
Normal 0.06477 0.04581 0.04337
RMSE
Extreme 0.07020 0.04663 0.04365
MAPE Normal 10.0892 7.39340 7.10707
(%) Extreme 11.8919 7.96620 7.49756
Normal 95.0187 96.3112 96.5049
Ap (%)
Extreme 94.7818 96.1243 96.6124
Normal 87.9875 91.0667 92.8333
As (%)
Extreme 87.1667 91.2000 92.7999

The robustness evaluation results of solar power as an

example are shown in Fig. 11. Compared to CGAN, WGAN-
GP and the proposed GAN have smoother discriminator
gradients, ensuring that they do not overfit noise components
and outliers in the input data, and maintain a relatively stable
generation. The robustness of WGAN-GP and the proposed
GAN is not significantly different, and their advantages
mainly lie in the optimized training parameters, which can
maintain relatively stable feedback during the training process,
so that the generator will not prematurely fall into an unstable
state during the optimization process.

Fig. 10. Accuracy of generated scenarios. VI. CONCLUSION

This paper considers the influence of extreme weather and
C. Comparison of Normal and Extreme Scenario Generation
presents an extreme scenario generation model based on
The significant difference between extreme and normal enhanced GAN. The proposed method has the potential
scenario generation is the training samples. The training applications in the field of extreme scenario generation. The
samples used in extreme scenario generation are characterized extreme scenario sample preparation process can achieve key
by many types and few quantities, while the training samples meteorological factor identification, weather-scenario
used in normal scenario generation show the opposite. The causality testing, as well as extreme scenario clustering and
comparison of the scenario generation is shown in TABLE IV. screening. The enhanced GAN model incorporates
Thanks to the advantages of training samples in normal meteorological labels and power system operation constraints,
scenarios, the matching quality of normal scenarios is better. reducing the dependence of the generated model on data and
However, there is little difference in the validity of the improving the quality of extreme scenario generation. The
generated samples. The proposed method has good work in sample preparation and generation model of the
adaptability to extreme scenario generation. proposed method can provide theoretical basis and technical
D. Robustness Analysis support for the extreme scenario generation in power systems.
Future work includes the following aspects. First, exploring
In the actual training process of the generation model, the
non-linear causal relationship analysis methods to evaluate the
samples often contain noise, outliers, etc., which may have
complex impact of meteorological factors. Second, optimize
adverse effects on the prediction results. The robustness
the efficiency and stability of model training. Explore new
evaluation of the model can find and solve potential problems

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 27,2025 at 16:06:06 UTC from IEEE Xplore. Restrictions apply.
© 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted,
but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Smart Grid. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSG.2025.3573166

> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <

loss functions to improve the training efficiency and stability [9] B. Qin, H. Wang, F. Li, D. Liu, Y. Liao, and H. Li, “Towards zero
carbon hydrogen: Co-production of photovoltaic electrolysis and natural
of the model. Finally, the adaptability of the generation model
gas reforming with CCS,” International Journal of Hydrogen Energy,
is expected to be tested in various complex scenarios. vol. 78, pp. 604–609, Aug. 2024, doi: 10.1016/j.ijhydene.2024.06.337.
[10] K. Wiel, L. P. Stoop, B. R. H. Zuijlen, R. Blackport, M. A. Broek, and F.
M. Selten, “Meteorological conditions leading to extreme low variable
renewable energy production and extreme high energy shortfall,”
Renewable and Sustainable Energy Reviews, vol. 111, pp. 261–275, Sep.
2019, doi: 10.1016/j.rser.2019.04.065.
[11] M. Ohba, Y. Kanno, and S. Bando, “Effects of meteorological and
climatological factors on extremely high residual load and possible
future changes,” Renewable and Sustainable Energy Reviews, vol. 175,
p. 113188, Apr. 2023, doi: 10.1016/j.rser.2023.113188.
[12] M. Amir, A. Amir, and G. Ali, “Weather sensitive short term load
forecasting using dynamic mode decomposition with control,” Electric
Power Systems Research, vol. 221, p. 109387, Aug. 2023, doi:
10.1016/j.epsr.2023.109387.
[13] J. Li, J. Zhou, and B. Chen, “Review of wind power scenario generation
methods for optimal operation of renewable energy systems,” Applied
Energy, vol. 280, p. 115992, Dec. 2020, doi:
10.1016/j.apenergy.2020.115992.
[14] H.-S. Li, X.-W. Wang, H. Nan, and M. Liu, “Application of a sampling-
based method for estimation of cumulative failure probability functions
of mechanisms,” Mechanism and Machine Theory, vol. 155, p. 104050,
Jan. 2021, doi: 10.1016/j.mechmachtheory.2020.104050.
[15] A. Tabandeh, G. Jia, and P. Gardoni, “A review and assessment of
importance sampling methods for reliability analysis,” Structural Safety,
vol. 97, p. 102216, Jul. 2022, doi: 10.1016/j.strusafe.2022.102216.
[16] P. Scarabaggio, S. Grammatico, R. Carli, and M. Dotoli, “Distributed
Demand Side Management With Stochastic Wind Power Forecasting,”
IEEE Transactions on Control Systems Technology, vol. 30, no. 1, pp.
97–112, Jan. 2022, doi: 10.1109/TCST.2021.3056751.
[17] P. Białas, P. Korcyl, and T. Stebel, “Analysis of autocorrelation times in
neural Markov chain Monte Carlo simulations,” Physical Review E, vol.
107, no. 1, p. 015303, Jan. 2023, doi: 10.1103/PhysRevE.107.015303.
[18] F. Zhang, L. Cheng, M. Wu, X. Xu, P. Wang, and Z. Liu, “Performance
analysis of two-stage thermoelectric generator model based on Latin
Fig. 11. Accuracy of points considering noise and outliers. hypercube sampling,” Energy Conversion and Management, vol. 221, p.
113159, Oct. 2020, doi: 10.1016/j.enconman.2020.113159.
REFERENCES [19] B. Qin, H. Wang, W. Li, F. Li, W. Wang, and T. Ding, “Aperiodic
Coordination Scheduling of Multiple PPLs in Shipboard Integrated
[1] B. Qin et al., “Challenges and opportunities for long-distance renewable Power Systems,” IEEE Transactions on Intelligent Transportation
energy transmission in China,” Sustainable Energy Technologies and Systems, pp. 1–11, 2024, doi: 10.1109/TITS.2024.3392832.
Assessments, vol. 69, p. 103925, Sep. 2024, doi: [20] R. Zhou, S. Chen, Y. Han, Q. Liu, Z. Chen, and W. Hu, “Probabilistic
10.1016/j.seta.2024.103925. spatiotemporal scenario generation method for dynamic optimal power
[2] A. T. D. Perera, V. M. Nik, D. Chen, J.-L. Scartezzini, and T. Hong, flow in distribution networks,” International Journal of Electrical
“Quantifying the impacts of climate change and extreme climate events Power & Energy Systems, vol. 155, p. 109667, Jan. 2024, doi:
on energy systems,” Nat Energy, vol. 5, no. 2, pp. 150–159, Feb. 2020, 10.1016/j.ijepes.2023.109667.
doi: 10.1038/s41560-020-0558-0. [21] A. B. Krishna and A. R. Abhyankar, “Time-coupled day-ahead wind
[3] H. Wang et al., “Optimal planning of hybrid hydrogen and battery power scenario generation: A combined regular vine copula and
energy storage for resilience enhancement using bi-layer decomposition variance reduction method,” Energy, vol. 265, p. 126173, Feb. 2023, doi:
algorithm,” Journal of Energy Storage, vol. 110, p. 115367, Feb. 2025, 10.1016/j.energy.2022.126173.
doi: 10.1016/j.est.2025.115367. [22] S. Park, Q. Xu, and B. F. Hobbs, “Comparing scenario reduction
[4] H. Li, B. Qin, S. Wang, T. Ding, J. Liu, and H. Wang, “Aggregate methods for stochastic transmission planning,” IET Generation,
power flexibility of multi-energy systems supported by dynamic Transmission & Distribution, vol. 13, no. 7, pp. 1005–1013, 2019, doi:
networks,” Applied Energy, vol. 377, p. 124565, Jan. 2025, doi: 10.1049/iet-gtd.2018.6362.
10.1016/j.apenergy.2024.124565. [23] W. Zhang and X. He, “A New Scenario Reduction Method Based on
[5] B. Qin, H. Wang, Y. Liao, D. Liu, Z. Wang, and F. Li, “Liquid Higher-Order Moments,” INFORMS Journal on Computing, vol. 34, no.
hydrogen superconducting transmission based super energy pipeline for 4, pp. 1903–1918, Jul. 2022, doi: 10.1287/ijoc.2021.1155.
Pacific Rim in the context of global energy sustainable development,” [24] W. Liao, Z. Yang, X. Chen, and Y. Li, “WindGMMN: Scenario
International Journal of Hydrogen Energy, vol. 56, pp. 1391–1396, Feb. Forecasting for Wind Power Using Generative Moment Matching
2024, doi: 10.1016/j.ijhydene.2023.12.289. Networks,” IEEE Transactions on Artificial Intelligence, vol. 3, no. 5,
[6] S. A. Shield, S. M. Quiring, J. V. Pino, and K. Buckstaff, “Major pp. 843–850, Oct. 2022, doi: 10.1109/TAI.2021.3128368.
impacts of weather events on the electrical power delivery system in the [25] I. Chopra and D. Selvamuthu, “Scenario generation in stochastic
United States,” Energy, vol. 218, p. 119434, Mar. 2021, doi: programming using principal component analysis based on moment-
10.1016/j.energy.2020.119434. matching approach,” OPSEARCH, vol. 57, no. 1, pp. 190–201, Mar.
[7] B. Qin et al., “New framework of low-carbon city development of 2020, doi: 10.1007/s12597-019-00418-8.
China: Underground space based integrated energy systems,” [26] J. Liang and W. Tang, “Scenario Reduction for Stochastic Day-Ahead
Underground Space, vol. 14, pp. 300–318, Feb. 2024, doi: Scheduling: A Mixed Autoencoder Based Time-Series Clustering
10.1016/j.undsp.2023.06.008. Approach,” IEEE Transactions on Smart Grid, vol. 12, no. 3, pp. 2652–
[8] P. L. Watson, A. Spaulding, M. Koukoula, and E. Anagnostou, 2662, May 2021, doi: 10.1109/TSG.2020.3047759.
“Improved quantitative prediction of power outages caused by extreme [27] H. Li et al., “A review of scenario analysis methods in planning and
weather events,” Weather and Climate Extremes, vol. 37, p. 100487, operation of modern power systems: Methodologies, applications, and
Sep. 2022, doi: 10.1016/j.wace.2022.100487.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 27,2025 at 16:06:06 UTC from IEEE Xplore. Restrictions apply.
© 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted,
but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Smart Grid. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSG.2025.3573166

> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <

challenges,” Electric Power Systems Research, vol. 205, p. 107722, Apr.

2022, doi: 10.1016/j.epsr.2021.107722.
[28] S. K. Mahjour, A. A. S. Santos, M. G. Correia, and D. J. Schiozer,
“Scenario reduction methodologies under uncertainties for reservoir
development purposes: distance-based clustering and metaheuristic
algorithm,” Journal of Petroleum Exploration and Production
Technology, vol. 11, no. 7, pp. 3079–3102, Jul. 2021, doi:
10.1007/s13202-021-01210-5.
[29] Q. Tu et al., “Forecasting Scenario Generation for Multiple Wind Farms
Considering Time-series Characteristics and Spatial-temporal
Correlation,” Journal of Modern Power Systems and Clean Energy, vol.
9, no. 4, pp. 837–848, Jul. 2021, doi: 10.35833/MPCE.2020.000935.
[30] P. W. Talbot et al., “Correlated synthetic time series generation for
energy system simulations using Fourier and ARMA signal processing,”
International Journal of Energy Research, vol. 44, no. 10, pp. 8144–
8155, 2020, doi: 10.1002/er.5115.
[31] L. Parviz and M. Ghorbanpour, “A hybrid EMD and MODWT models
for monthly precipitation forecasting using an innovative error
decomposition method,” Stochastic Environmental Research and Risk
Assessment, Aug. 2024, doi: 10.1007/s00477-024-02797-x.
[32] R. Luzia, L. Rubio, and C. E. Velasquez, “Sensitivity analysis for
forecasting Brazilian electricity demand using artificial neural networks
and hybrid models based on Autoregressive Integrated Moving
Average,” Energy, vol. 274, p. 127365, Jul. 2023, doi:
10.1016/j.energy.2023.127365.
[33] S. Bhavsar, R. Pitchumani, and M. A. Ortega-Vazquez, “Machine
learning enabled reduced-order scenario generation for stochastic
analysis of solar power forecasts,” Applied Energy, vol. 293, p. 116964,
Jul. 2021, doi: 10.1016/j.apenergy.2021.116964.
[34] B. Ivanovic, K. Leung, E. Schmerling, and M. Pavone, “Multimodal
Deep Generative Models for Trajectory Prediction: A Conditional
Variational Autoencoder Approach,” IEEE Robotics and Automation
Letters, vol. 6, no. 2, pp. 295–302, Apr. 2021, doi:
10.1109/LRA.2020.3043163.
[35] Y. Chen, Y. Wang, D. Kirschen, and B. Zhang, “Model-Free Renewable
Scenario Generation Using Generative Adversarial Networks,” IEEE
Transactions on Power Systems, vol. 33, no. 3, pp. 3265–3275, May
2018, doi: 10.1109/TPWRS.2018.2794541.
[36] Y. Akkem, S. K. Biswas, and A. Varanasi, “A comprehensive review of
synthetic data generation in smart farming by using variational
autoencoder and generative adversarial network,” Engineering
Applications of Artificial Intelligence, vol. 131, p. 107881, May 2024,
doi: 10.1016/j.engappai.2024.107881.
[37] J. Gui, Z. Sun, Y. Wen, D. Tao, and J. Ye, “A Review on Generative
Adversarial Networks: Algorithms, Theory, and Applications,” IEEE
Transactions on Knowledge and Data Engineering, vol. 35, no. 4, pp.
3313–3332, Apr. 2023, doi: 10.1109/TKDE.2021.3130191.
[38] W. Dong, X. Chen, and Q. Yang, “Data-driven scenario generation of
renewable energy production based on controllable generative
adversarial networks with interpretability,” Applied Energy, vol. 308, p.
118387, Feb. 2022, doi: 10.1016/j.apenergy.2021.118387.
[39] Y. Zhang, Q. Ai, F. Xiao, R. Hao, and T. Lu, “Typical wind power
scenario generation for multiple wind farms using conditional improved
Wasserstein generative adversarial network,” International Journal of
Electrical Power & Energy Systems, vol. 114, p. 105388, Jan. 2020, doi:
10.1016/j.ijepes.2019.105388.
[40] R. Yuan, B. Wang, Y. Sun, X. Song, and J. Watada, “Conditional Style-
Based Generative Adversarial Networks for Renewable Scenario
Generation,” IEEE Transactions on Power Systems, vol. 38, no. 2, pp.
1281–1296, Mar. 2023, doi: 10.1109/TPWRS.2022.3170992.
[41] Y. Li, J. Li, and Y. Wang, “Privacy-Preserving Spatiotemporal Scenario
Generation of Renewable Energies: A Federated Deep Generative
Learning Approach,” IEEE Transactions on Industrial Informatics, vol.
18, no. 4, pp. 2310–2320, Apr. 2022, doi: 10.1109/TII.2021.3098259.
[42] L. Yin and C. Lin, “Matrix Wasserstein distance generative adversarial
network with gradient penalty for fast low-carbon economic dispatch of
novel power systems,” Energy, vol. 298, p. 131357, Jul. 2024, doi:
10.1016/j.energy.2024.131357.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 27,2025 at 16:06:06 UTC from IEEE Xplore. Restrictions apply.
© 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted,
but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information.