Hindawi

Complexity
Volume 2019, Article ID 1534598, 23 pages
https://doi.org/10.1155/2019/1534598

Research Article
Planning of Cascade Hydropower Stations with the
Consideration of Long-Term Operations under Uncertainties

1 1,2
Changjun Wang and Shutong Chen
1
Glorious Sun School of Business and Management, Donghua University, Shanghai 200051, China
2
Department of Industrial and Systems Engineering, Virginia Tech, Blacksburg, VA 24061, USA

Correspondence should be addressed to Changjun Wang; [email protected]

Received 8 July 2019; Revised 21 September 2019; Accepted 8 October 2019; Published 28 November 2019

Guest Editor: Chun Wei

Copyright © 2019 Changjun Wang and Shutong Chen. This is an open access article distributed under the Creative Commons
Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.
In the location-related planning of a hydropower system, the consideration of future operations under uncertainties can make
the decisions sustainable and robust. Then, it is of great importance to develop an effective approach that deals with the long-
term stochasticity due to the long-lasting effects of the location selections. Thus, we propose a multistage stochastic pro-
gramming model to optimize the planning decisions of cascade hydropower stations and the long-term stochastic operations in
an integrated way. The first stage (i.e., the planning stage) in the model deals with the location and capacity decisions of the
hydropower stations, while the subsequent stages implement the scheduling decisions under each stagewise stochastic scenario.
To address the curse of dimensionality caused by the long-term stochastic operations, we further propose a novel di-
mensionality reduction approach based on dual equilibrium to transform the multistage model into a tractable two-stage
stochastic program. The applicability of our approach is validated by a case study based on a basin of Yangtze River, China, and
corresponding sensitivity analysis.

1. Introduction is far beyond the long-term scale considered in previous

works, such as 3–5 years in Gjelsvik et al. [4] or 12-period
Given a basin, the hydropower stations to-be-built usually used in Cheng et al. [5], Wang et al. [6], and Xu and Mei [7],
could be multiple. Even only a hydropower station in the because of the long-lasting effects of the location decisions.
current timetable, it is still necessary to consider the impacts Hence, the challenge is how to handle such overlong sto-
of other possible hydropower stations in the future because chastic nature.
of their interdependence. Moreover, such a plan that in- In previous works, the decisions on locations and
volves major capital investments must be made for a operations of a hydropower system have been extensively,
longstanding future. Hence, subsequent operations of all the but separately, studied. Specifically, when a decision-
stations, which influence the effect of the planning decisions, maker plans a hydropower station, advanced technolo-
should be considered. Such an integrated way can help to get gies, such as geographical information system (GIS), are
the global optimum in disasters mitigation, irrigation water often utilized to obtain geological, climatic, and topo-
supply, and clean power production [1, 2]. However, in the graphical information in most recent studies. Based on
long-term operations, natural or social uncertainties, such as this information, different mathematical methods are
water inflows and demands, are inevitable and influence the further utilized for spotting site by synthesizing natural,
performance of the hydropower system significantly [3]. engineering, and social criteria. For example, Larentis
Thus, the problem faced by the decision-maker is long-term et al. [8] propose a GIS-based DEM (digital elevation
stochastic optimization. The “long-term” we mentioned here model) composed of the preidentification of promising
2 Complexity

sites and the multicriteria feasibility assessment of the Uncertainties extensively exist in hydropower systems.
final set, in which energy, technical, and environmental For example, water inflows usually vary and cannot be
factors are considered together. Another GIS-based DEM accurately predicted since long-term meteorological fore-
is proposed by Kusre et al. [9] which use a hydrologic casts are unreliable [29, 30]. Thus, the optimization of
model to assess water resource utilization under different hydropower operations under uncertainties is basically a
site candidates. Similar works can be referred to Serpoush risk-based decision-making problem [31]. The stagewise
et al. [10], Zaidi and Khan [11], etc. Besides the above stochastic process can be represented as a scenario tree;
assessment models, the optimization technique becomes hence, the multistage stochastic programming (MSSP)
another alternative in recent years. For example, Hosnar model, which optimizes the expected value based on the
and Kovač-Kralj [12] identify the optimal installation scenario set, has been popularly used. For example, Fleten
locations by maximizing an ecoprofit objective at tech- and Kristoffersen [32] develop an MSSP model to make the
nological, economic, environmental, and social con- decisions on power generation, in which uncertain water
straints. To identify the appropriate hydropower dam inflows and electricity market prices are considered, with
location, Loannidou and O’Hanley [13] develop a mixed- the optimization of the expected benefit. Chazarra et al.
integer linear programming model to optimize the hy- [33] propose an MSSP model for a hydropower system,
dropower potential with the consideration of river con- taking uncertain water inflows and electricity market prices
nectivity. In these studies, the future power generation into account, to simultaneously maximize the expected
and hydrologic dynamics are described by the empirical profit in both energy and regulation reserve markets. To
formulas in the cumulative form. Thus, time-varying deal with uncertain streamflow, a multiobjective MSSP
operations under uncertainties are ignored in the setting model is presented by Xu et al. [34] to optimize scheduling
of location-related planning. strategies. The first objective is to maximize direct revenue
The optimization of the multistage operational ac- from energy production, and the second one is to minimize
tivities, such as water storage, supply, and power gen- the expected energy shortfall percentage. Séguin et al. [35]
eration, is an important issue which has drawn lots of address the stochastic hydropower unit commitment and
studies. Most of them focus on a given hydropower loading problem under the uncertain inflows. However, all
system involving a single hydropower plant (e.g., Vieira these works only consider the short-term horizon. Thus,
et al. [14]) or multistations. For the multistation oper- off-the-shelf solvers can be applied to solve these models
ations in the deterministic setting, two main stream of directly.
approaches have been applied. At first, to pursue the With the increase of the number of decision stages, the
optimality, mathematical programming methods, espe- size of the MSSP grows dramatically which often requires
cially dynamic programming (DP), are often used but decomposition methods [36]. A state-of-the-art one is the
also suffer from the computational challenge due to the stochastic dual DP algorithm which is a sampling-based
number of stages and stations. Thus, Cheng et al. [15] variant of nested Benders decomposition. Hjelmeland et al.
reduce the scale of the problem by limiting their work in [37] use it to handle a medium-term scheduling issue for a
the “short-term” horizon. Li et al. [16] consider a de- single producer under uncertain inflows and prices within
composition-coordination mechanism to reduce di- one to three-year horizon. Most related works consider the
mensionality. Li et al. [17] and Cheng et al. [5] propose to medium-term setting as Gjelsvik et al. [4]; Helseth et al. [38];
parallelize the DP algorithm to reduce the computation Hjelmeland et al. [39]; and Poorsepahy-Samian et al. [40].
time. The former uses the distributed memory archi- However, as indicated by Hjelmeland et al. [39], although
tecture and the message passing interface protocol, while decomposition methods can help to alleviate the solving
the latter considers the Fork/Join parallel framework in a complexity, the computation would significantly become
multicore environment. Feng et al. [18] and Feng et al. slow with the increase of system size and decision stages. It
[19] focus on the simplification of state set to accelerate shows the difficulty of multistage stochastic optimization
the implementation of DP. Cheng et al. [20] and Feng with high dimensions of uncertainty.
et al. [21] adopt progressive optimality algorithm to In summary, a considerable amount of research has
modify the conventional DP by dividing the multistage been conducted for the hydropower system design and
problem into a sequence of subproblems and thus reduce operation. However, to the best of our knowledge, the
the computational burden. Second, in order to address location decisions in existing studies have not taken
the computational complexity, numerous heuristic al- multistage operations into account. In terms of operations,
gorithms have been considered by sacrificing some op- numerous studies focus on developing efficient exact or
timality in recent years. Typical methods used in this heuristic algorithms for the deterministic setting, whereas
stream include particle swarm optimization [22, 23, 24], many endeavors utilize the MSSP to address the stochastic
electromagnetism-like mechanism [25, 26], genetic al- operations. However, the high dimensionality character-
gorithm [27], water cycle algorithm [7], and artificial istic has become the main bottleneck limiting the appli-
intelligence algorithms [28]. However, all these studies cation of optimization approaches [41]. Even in the
focus on the deterministic setting. deterministic setting, the computation of the hydropower
Complexity 3

Outside water Outside water Station i + 1

inflow i Station i inflow i + 1

Water discharge i Capacity Water discharge i + 1

...
Capacity i+1 Water storage ...
i Water storage Water for power Abandoned i+1
i generation i water i

Loss flow i Loss flow i + 1

Water demand i Water demand i + 1

Figure 1: Illustration of operational activities.

system operation problem is highly complex when the 2. Problem Definition and
number of operational stages is big [19]. With the con- Mathematical Formulation
sideration of randomness, the optimization of the MSSP
should be implemented on all the stagewise stochastic 2.1. Problem Statement. Consider a basin, where a govern-
realizations, which further increase the solving burden. ment prepares a construction plan of cascade hydropower
Hence, most existing works only focus on short- or me- stations to control flood or drought, satisfy irrigation, and
dium-term stochasticity. However, the location selection is pursue profits from power generation. Thus, the location and
the planning decision that involves major capital in- capacity decision-making problem of cascade hydropower
vestments and has an overlong effect. Thus, such a decision stations is studied here, in which the long-term stochastic
should be made from the long-lasting perspective, in which operations should be taken into account. The integrated
uncertainties would occur inevitably. Hence, the long-term MSSP model is formulated to optimize the total performance
stochasticity should be considered in the hydropower which involves the construction costs in the planning stage, as
design, which is a difficult task. well as operational costs, penalty costs, and profits of power
We contribute to the existing literature by planning a generation in the subsequent multiple stages.
cascade hydropower system with the consideration of Six decisions of the model can be divided into two
long-term stochastic operations. Most mainstream algo- groups. The first group should be made in the planning stage,
rithms, including DP for the deterministic setting and the including (1) the final selection of hydropower stations from
decomposition methods for the stochastic setting, suffer the candidates; (2) the capacity of each selected hydropower
from the number of stages. Few papers have looked spe- station. The decisions in the second group are in each op-
cifically into long-term stochastic models. Considering erational stage. They are (3) water storage of each selected
that the purpose of our work is the location decisions hydropower station in each stage; (4) loss flow of each se-
rather than providing accurate schedules, in order to lected hydropower station in each stage; (5) water discharge
address the dimensionality issue, an intuitive idea is to to downstream from each hydropower station in each stage;
keep the influence of stochastic operations while reducing and (6) water discharge of power generation in each selected
the number of stages. Thus, applying the method of hydropower station each stage. During these operational
dual equilibrium (DE) [42, 43], we propose a novel di- stages, we consider uncertain water inflows (including
mensionality reduction approach which aggregates the precipitation and inflows from tributaries) and water de-
long-term operational impacts on the present and thus mands (from residents, agriculture, and industry). The
simplifies the MSSP model to a two-stage one. Our method number of stages, defined as T, could be extremely large.
can handle the stochastic problem regardless of the As shown in Figure 1, given the locations of cascade
number of stages and provide an alternative approach for hydropower stations, three scheduling activities: water
the overlong setting. discharge of power generation, abandoned water spill (go
The remainder of the study is presented as follows. In downstream directly without passing generator units) and
Section 2, the MSSP model is developed. The di- loss flow should be periodically made. The water storage of
mensionality reduction approach based on DE is presented each hydropower station is determined by its scheduling
in Section 3. Specifically, Section 3.1 focuses on the re- activities and random water inflows. Notice that the loss flow
duction of the scenario tree while Section 3.2 implements mentioned is discharged to local areas, rather than down-
the transformation of the MSSP model accordingly. In stream. When the loss flow is larger than the actual demand,
Section 4, the application in one section of Yangtze River, it would incur corresponding penalties. Moreover, the
China, is displayed and analyzed. The last section presents physical limits on the levels of water discharge, loss flow, and
conclusions and remarks about some directions for future power generation of each hydropower station should be
research. satisfied in each operational stage.
4 Complexity

Table 1: Model variables.

Types of variables Symbol Description
R Set of hydropower station candidates
Sets
S Set of scenarios of the operational stages
The planning horizon is composed of T operational
T
stages
The minimum required number of hydropower
Rmin
stations in the plan
e Unit construction fee (RMB/m3)
f Unit profit of power generation (RMB/KWh)
Time interval of each operational stage (h, which is
Δt
the abbreviation of hour)
Penalty coefficient per unit in hydropower station
ci
candidate i (RMB/m3)
Ki Coefficient of output of candidate i (kg/(m2·s2))
Deterministic variables
Nimin The minimum output of candidate i (kW)
Nimax The maximum output of candidate i (kW)
Hi Water head of candidate i (m)
The minimum required water storage of candidate i
Vimin
when it is selected (m3)
The maximum discharge limit of local area around
Qimax
candidate i (m3/h)
QFimax The upper limit water discharge of candidate i (m3/h)
QFimin The lower limit water discharge of candidate i (m3/h)
Unit operational cost of candidate i at stage t (RMB/
ri(t)
(m3·h))
ps The probability of stagewise scenario s
Outside water inflows in candidate i at stage t under
εsi (t)
scenario s (m3/h)
Water demands in candidate i at stage t under
Stochastic variables Qsi min (t)
scenario s (m3/h)
Acceptance level of water discharge in local area
Q′simin (t) surrounding candidate i at stage t under scenario s
(m3/h)

Table 2: Decision variables.

Types of variables Symbol Description
xi 1 if hydropower station i is chosen, 0 otherwise
The planning stage The maximum capacity of hydropower station
Capi
candidate i
Water storage of candidate i at stage t under scenario
ysi (t)
s (m3)
Loss flow of candidate i at stage t under scenario s
QLsi (t)
(m3/h)
Water discharge to downstream in candidate i at
The operational stages Qsi (t)
stage t under scenario s (m3/h)
Water discharge of power generation in candidate i at
QEsi (t)
stage t under scenario s (m3/h)
Water spill in candidate i at stage t under scenario s
QAsi (t)
(m3/h)

2.2. Model Formulation. Sets, exogenous deterministic and The MSSP model which integrates the planning de-
stochastic variables are given in Table 1. Table 2 displays the cisions and stochastic operational impacts is proposed as
decision variables in all stages. follows:
Complexity 5

T
min ⎝Δt · 􏽘 r (t) · Cap ⎞
􏽘 e · Capi + 􏽘⎛ ⎠
i i
i∈R t�1 i∈R
(1a)
T
+ 􏽘 􏽘⎡ s
⎣p · 􏽘􏼔Δt · 􏼒ci · xi · s
􏼒QLi (t) − Q′simin (t)􏼓 − f · Ki · QEsi (t) · Hi 􏼓􏼕⎤⎦,
t�1 s∈S i∈R

s.t.
Rmin ≤ 􏽘 xi ≤ R, (1b)
i∈R

Capi ≤ MM · xi , i ∈ R, (1c)

QEsi (t) ≤ MM · xi , t � 1, 2, . . . , T; i ∈ R; s ∈ S, (1d)

xi · Ni min ≤ Ki · QEsi (t) · Hi ≤ Ni max , t � 1, 2, . . . , T; i ∈ R; s ∈ S, (1e)

Vi min · xi ≤ ysi (t) ≤ Capi , t � 1, 2, . . . , T; i ∈ R; s ∈ S, (1f)

ysi (t) − ysi (t − 1) � Δt · 􏼂Qsi− 1 (t) + εsi (t) − Qsi (t) − QLsi (t)􏼃, t � 1, 2, . . . , T; i ∈ R; s ∈ S, (1g)

Qsi (t) � QEsi (t) + QAsi (t), t � 1, 2, . . . , T; i ∈ R; s ∈ S, (1h)

QFi min ≤ QEsi (t) + QAsi (t) ≤ QFi max , t � 1, 2, . . . , T; i ∈ R; s ∈ S, (1i)

Qsi min (t) ≤ QLsi (t) ≤ xi · Qi max + 1 − xi 􏼁 · Qsi min (t), t � 1, 2, . . . , T; i ∈ R; s ∈ S, (1j)

Capi , ysi (t), QLsi (t), QEsi (t), Qsi (t), QAsi (t) ≥ 0, xi ∈ {0, 1}, t � 1, 2, . . . , T; i ∈ R; s ∈ S. (1k)

The objective function (1a) is composed of the con- 3. Model Transformation

struction costs of all selected hydropower stations and the
expected value of the operational costs, the environmental Even the deterministic version of the proposed MSSP
penalty costs, and the benefits of power generation under all model is NP-hard. Besides, the dimension of stochastic
the stagewise scenarios. scenarios in the MSSP model increases with the number of
Constraint (1b) limits the number of hydropower sta- stages exponentially. It would further increase the solving
tions to be chosen. Thus, Constraints (1c) and (1d) ensure difficulty. DE is a useful approach to handle the dimension
that the corresponding maximum capacity and water dis- problem of the large-scale setting [44]. It was first proposed
charge of unselected hydropower stations are zero in which in [42] to simplify the multistage deterministic convex
MM is a number big enough. Constraint (1e) gives the power optimization problem and then was applied to a multistage
generation limits. Constraint (1f ) ensures that the water stochastic production-inventory programming model in
storage of the selected hydropower station is between the [43]. The essence of DE is going to add up all the influence
minimum required water storage and maximum capacity. after a time point in the future by a so-called discount factor
Constraint (1g) gives water balance equations. Constraint to compress stages. Hence, it coincides with our idea of
(1h) shows the relationship between water discharge to aggregating the operational impacts on the planning
downstream and water discharge of power generation, as decisions.
well as water spill in each hydropower station candidate. Next, we proposed a DE-based dimensionality reduction
Constraint (1i) presents the upper and lower limits of water approach to address the above MSSP model. Specifically,
discharge between two adjacent selected hydropower sta- because the size of the scenario tree is one of the main causes
tions. Constraint (1j) guarantees that if the candidate is resulting in the computational difficulty, we first show how
chosen, its loss flow should meet the local demand and to compress the stages of the scenario tree at first (see Section
cannot exceed the maximum discharge limit. And if a 3.1). Along with the simplification of the scenario tree, we
candidate is not selected, its loss flow is equal to the water further show how to aggregate the operational decision
demands. Constraint (1k) specifies the domains of the de- variables and the deterministic parameters. With the
cision variables. compression of all these variables, the proposed MSSP
6 Complexity

model would be integrated into a two-stage one which is Figure 4). Specifically, take the inflows of candidate 1 in the
tractable (see Section 3.2). first stage (t � 1), i.e., εξ1 (1), as an example. Without loss of
generality, assume that T is the multiple of four. The
subsequent water inflows with the same season, i.e., stage
3.1. Scenarios Generation and Simplification. Because the 4τ + 1 (τ �1, 2, . . ., (T/4)–1), have the same value, denoted
acceptance level of water discharge includes the quantity as ε1 (1). Similarly, ε1 (2), ε1 (3), and ε1 (4) represent other
of water demands, we assume these two factors have a three seasonal values of water inflows at candidate 1.
deterministic linear relationship. Thus, we focus on Referring to the concept of present value, we use factor δ
random water inflows and water demands here. Before we (0 < δ < 1) to discount the impacts of subsequent stages.
give the dimensionality reduction approach, we first show Thus, the corresponding aggregated water inflows of
the generation and simplification procedures of sto- candidate 1 in the first stage (t � 1) under the case of ξ,
chastic scenarios by taking uncertain water inflows as an denoted as ε∗ξ1 (1), are the discounted value which can be
example. Uncertain water demands can be handled expressed as
similarly.
We match each operational stage with a season. At ε∗ξ ξ 4 4τ
1 (1) � ε1 (1) + δ · ε1 (1) + · · · + δ · ε1 (1) + · · · + δ
T− 4
· ε1 (1)
seasonal stage t � 1, 2, . . ., T, assume that ξ � 1, 2, and 3
correspond to three levels of random inflows: high, medium, 1 − δT T
� εξ1 (1) + ε1 (1) · , τ � 1, 2, . . . , − 1.
and low, with the corresponding probability of p1H, p1M, and 1 − δ4 4
p1L, respectively. The quantities of water inflows in hydro- (2)
power station candidate i at stage t with level ξ can be
represented as εξi (t), i ∈ R. There are multiple stages con- Then, ε∗ξ ∗ξ ∗ξ
1 (2), ε1 (3), and ε1 (4) can be obtained simi-
sidered in this study. Hence, each stagewise scenario of larly. Thus, the scenario tree in Figure 3 is compressed into
inflows, defined as s1 (∈S1), can be represented as a branch Figure 5, in which the value associated with each node is a
from t � 0∼T of the scenario tree, in which t � 0 represents discounted value now. However, the number of branches
the planning stage. The corresponding probability p(s1) is the (􏽢s1 ) and their corresponding occurrence probabilities (p(􏽢s1 ))
product of occurrence probabilities of all the nodes on this remain unchanged.
branch. Figure 2 gives the scenario tree faced by hydropower
station candidate 1.
As can be seen, the number of inflow scenarios of one 3.1.2. Step 2 of Dimensionality Reduction—Stages Reduction
candidate is up to 3T in T seasonal stages. For example, if Across Seasons. It is worth recalling that this study focuses
we consider a 10 year horizon, the scenario size of inflows on the planning problem with the consideration of future
of one candidate would be 340(≈1.049 × 1019). Obviously, a operational activities under uncertainties. Thus, the second
simplification way is required. Here, we can remain the step of the dimensionality reduction approach is to im-
scenarios in early stages, while inflows in subsequent plement the further aggregation of the impacts of opera-
stages are replaced by corresponding constant estimated tional stages together. In other words, the random variables
values. We do this for two reasons. The first is that the (t � 2, 3, 4) on a branch in Figure 5 will be integrated into the
stochastic scenarios in subsequent stages are harder to give first operational stage (t � 1). For the example that the in-
than these of early stages due to forecasting difficulty. flows of candidate 1 under stagewise scenario 􏽢s1 which is
Second, random inflows have intrinsic periodicity which composed of four high-level stages (i.e., ξ � 1), its present
means inflows in subsequent stages can be offset. Hence, value can be given as
an approximated way is to use the corresponding esti- ∗∗􏽢
s
ε1 1 � ε∗1 ∗1 2 ∗1 3 ∗1
1 (1) + δ · ε1 (2) + δ · ε1 (3) + δ · ε1 (4).
(3)
mated values to replace the seasonal inflows in subsequent
stages. Although the impacts of scenario 􏽢s1 is aggregated now,
Thus, the scenario tree in Figure 2 is simplified to that in such process would not change the number of scenario 􏽢s1
Figure 3 in which the uncertainties of t � 1∼4 remain. Denote and its probability p(􏽢s1 ).
the stagewise scenario in Figure 3 as 􏽢s1 and the corre- The stochastic scenario of random water demands could
sponding scenario set as 􏽢S1 . After simplification, the number be handled similarly, and the corresponding final scenario
of the decision variables is still large because of the big T. set is denoted as 􏽢S2 which is composed by scenario 􏽢s2 with
Hence, the dimensionality reduction approach is proposed the probability p(􏽢s2 ). Hence, the set of total scenarios S used
to address this issue. in the proposed MSSP model is simplified to 􏽢S, which is the
Cartesian product of 􏽢S1 and 􏽢S2 :
􏼌
3.1.1. Step 1 of Dimensionality Reduction—Stages Reduction 􏽢S � 􏽢S1 × 􏽢S2 � 􏽮 􏽢s1 , 􏽢s2 􏼁 􏼌􏼌􏼌 􏽢s1 ∈ 􏽢S1 , 􏽢s2 ∈ 􏽢S2 􏽯. (4)
by Aggregating the Random Factors of the Same Season.
The dimensionality reduction approach includes two steps. The corresponding probability of new scenario 􏽢s (∈􏽢S),
We focus on the first one in this section. which is the union of 􏽢s1 and 􏽢s2 , can be calculated as
Due to the seasonal periodicity of inflows, we propose
to aggregate the inflows of the same season together (see p(􏽢s) � p 􏽢s1 􏼁 × p 􏽢s2 􏼁. (5)
Complexity 7

t=0 t=1 t=2 t=3 t=4 t=5 ... t=T

...

... ...

... ...

Candidate 1 ...
... ...

...
Figure 2: Original multistage scenario tree (random inflows of candidate 1).

t=0 t=1 t=2 t=3 t=4 t=5 ... t=T

...

...

Candidate 1 ...

...
Constant predicted value
Figure 3: Simplified multistage scenario tree (random inflows of candidate 1).

The first year The second year The τ – 1 year

...

t=0 t=1 t=2 t=3 t=4 t=5 t=6 t=7 t=8 ... t= t= t= t= ... t=T
4τ + 1 4τ + 2 4τ + 3 4τ + 4

Figure 4: Illustration of stage reduction by aggregating the random factors of the same season.

Till now, the stagewise scenario tree is compressed into a variables. Accordingly, with the compression of the scenario
one-stage scenario set 􏽢S. tree, the deterministic parameters and the decision variables
should also be aggregated, and thus, the MSSP model would
3.2. Transformation of the MSSP Model. The above work be simplified. The transformation process also includes two
simplifies the scenario tree by discounting the random steps.
8 Complexity

t=4

t=3 ε∗1
1 (4)

1H
p
t=2 ε∗1 p1M ε∗2
1 (4)
1 (3)

p 1L
1H
p
p1M ε∗2 ε∗3
1 (4)
t=1 ε∗1
1 (2) 1 (3)
...

p 1L
1H
p
ε∗1 p1M ε∗2 ... ε∗3
1 (3)
...
t=0 1 (1) 1 (2)

p 1L
1H
p

ε∗3
1 (2)
...
p1M Part of
Candidate 1 ε∗2
1 (1)
...
Set S1
ε∗1
1 (2)
...
p 1L

1H
p

ε∗3
1 (1) p1M ε∗2 ... ε∗1
1 (3) ...
1 (2)
p 1L

1H
p

ε∗3 p1M ... ε∗1

1 (4)
1 (2) ε∗2
1 (3)

1H
p 1L

p
ε∗3
1 (3) p1M ε∗2
1 (4)
p 1L

ε∗3
1 (4)

Figure 5: Scenario tree (random inflows of candidate 1).

3.2.1. Step 1 of Dimensionality Reduction. At first, we de- scenario. For the submodel, we show how to simplify the
compose the operational stages of MSSP model into variables and the model along with the reduction of the
multiple submodels according to each stagewise scenario s stages.
(∈S). Then, such submodel is not only multistage but also To facilitate the implementation of DE, we first refor-
deterministic because it is given based on the independent mulate the submodel as the following equivalent:

4 (T/4)− 1⎧
⎨ ⎬
⎫
min 􏽘 􏽘 ⎩ 􏽘 Δt · 􏼂ci · xi · QLi (4τ + j) − Q′i min(4τ + j)􏼁 − f · Ki · QEi (4τ + j) · Hi 􏼃⎭ , (6a)
j�1 τ�0 i∈R

s.t.
T (6b)
QEi (4τ + j) ≤ MM · xi , i ∈ R; τ � 0, . . . , − 1; j � 1, 2, 3, 4,
4
T
xi · Ni min ≤ Ki · QEi (4τ + j) · Hi ≤ Ni max , i ∈ R; τ � 0, . . . , − 1; j � 1, 2, 3, 4, (6c)
4
T
Vi min · xi ≤ yi (4τ + j) ≤ Capi , i ∈ R; τ � 0, . . . , − 1; j � 1, 2, 3, 4, (6d)
4
Complexity 9

yi (4τ + j) − yi (4τ + j − 1) � Δt · 􏼂Qi− 1 (4τ + j) + εi (4τ + j) − Qi (4τ + j) − QLi (4τ + j)􏼃,

T (6e)
i ∈ R; τ � 0, . . . , − 1; j � 1, 2, 3, 4,
4
T
Qi (4τ + j) � QEi (4τ + j) + QAi (4τ + j), i ∈ R; τ � 0, . . . , − 1; j � 1, 2, 3, 4, (6f)
4
T
QFi min ≤ QEi (4τ + j) + QAi (4τ + j) ≤ QFi max , i ∈ R; τ � 0, . . . , − 1; j � 1, 2, 3, 4, (6g)
4

Qi min (4τ + j) ≤ QLi (4τ + j) ≤ xi · Qi max + 1 − xi 􏼁 · Qi min (4τ + j),

T (6h)
i ∈ R; τ � 0, . . . , − 1; j � 1, 2, 3, 4.
4

Step 1. Here, we focus on the objective. Specifically, compressed into the corresponding stage j. Thus, with the
similar to formula (2), the objective values at stage 4τ + j introduction of the discount factor δ, (6a) can be
(τ � 0, 1, . . ., (T/4)–1, j � 1, 2, 3, 4) also should be reformulated as (7):

4 (T/4)− 1 ⎧
⎨ ⎬
⎫
min 􏽘 􏽘 ⎩ δ4τ · 􏽘 Δt · 􏼂ci · xi · QLi (4τ + j) − Qi′min(4τ + j)􏼁 − f · Ki · QEi (4τ + j) · Hi 􏼃⎭ . (7)
j�1 τ�0 i∈R

Step 2. Along with simplification of the scenario tree, the use dual multipliers βk,4τ+j (k � 1, . . ., 7) to relax the k-th one
decision variables and the deterministic parameters should of Constraints (6b)∼(6h) of stage 4τ + j (τ � 0, 1, . . ., (T/4)–1,
also be aggregated. Because they exist in both the objective j � 1, 2, 3, 4). Then, the optimization problem ((6b)∼(6h),
and the constraints, to facilitate our transformation, we (7)) can be transformed into the unconstrainted problem
integrate the constraints into the objective. Specifically, we (8):

4 (T/4)− 1
min 􏽘 􏽘 􏼨δ4τ · 􏽘 Δt · 􏼂ci · xi · QLi (4τ + j) − Qi′min(4τ + j)􏼁 − f · Ki · QEi (4τ + j) · Hi 􏼃
j�1 τ�0 i∈R

+ β+1,4τ+j · QEi (4τ + j) − MM · xi 􏼁 + β+2,4τ+j · Ki · QEi (4τ + j) · Hi − Ni max 􏼁 + β−2,4τ+j

· Ki · QEi (4τ + j) · Hi − xi · Ni min 􏼁 + β+3,4τ+j · yi (4τ + j) − Capi 􏼁 + β−3,4τ+j (j) · yi (4τ + j) − Vi min · xi 􏼁
+ β4,4τ+j · yi (4τ + j) − yi (4τ + j − 1) − Δt · Qi− 1 (4τ + j) + εi (4τ + j) − Qi (4τ + j) − QLi (4τ + j)􏼁􏼁 (8)

+ β5,4τ+j · Qi (4τ + j) − QEi (4τ + j) − QAi (4τ + j)􏼁 + β+6,4τ+j · QEi (4τ + j) + QAi (4τ + j) − QFi max 􏼁
+ β−6,4τ+j · QEi (4τ + j) + QAi (4τ + j) − QFi min 􏼁 + β+7,4τ+j · QLi (4τ + j) − xi · Qi max − 1 − xi 􏼁 · Qi min (4τ + j)􏼁

+ β−7,4τ+j · QLi (4τ + j) − Qi min (4τ + j)􏼁􏼩.

The DE approach assumes the dual multipliers have the the dual multipliers here take the form βk,4τ+j � δ4τ·βk(j).
linear relationship with the discount factor δ [42]. Thus, let Then, problem (8) can further be reformulated as
10 Complexity

4 (T/4)− 1
min 􏽘 􏽘 δ4τ · 􏼨 􏽘 Δt · 􏼂ci · xi · QLi (4τ + j) − Qi min′(4τ + j)􏼁 − f · Ki · QEi (4τ + j) · Hi 􏼃 + β+1 (j) · QEi (4τ + j) − MM · xi 􏼁
j�1 τ�0 i∈R

+ β+2 (j) · Ki · QEi (4τ + j) · Hi − Ni max 􏼁 + β−2 (j) · Ki · QEi (4τ + j) · Hi − xi · Ni min 􏼁 + β+3 (j) · yi (4τ + j) − Capi 􏼁

+ β−3 (j) · yi (4τ + j) − Vi min · xi 􏼁 + β4 (j) · 􏼠yi (4τ + j) − yi (4τ + j − 1) − Δt · 􏼠Qi− 1 (4τ + j) + εi (4τ + j)

− Qi (4τ + j) − QLi (4τ + j)􏼡􏼡 + β5 (j) · Qi (4τ + j) − QEi (4τ + j) − QAi (4τ + j)􏼁 + β+6 (j) · 􏼠QEi (4τ + j)

+ QAi (4τ + j) − QFi max 􏼡 + β−6 (j) · QEi (4τ + j) + QAi (4τ + j) − QFi min 􏼁

+ β+7 (j) · QLi (4τ + j) − xi · Qi max − 1 − xi 􏼁 · Qi min (4τ + j)􏼁 + β−7 (j) · QLi (4τ + j) − Qi min (4τ + j)􏼁􏼩.

(9)

Step 3. Thus, similar to the discounting way of the random

(T/4)− 1
variables in Section 3.1.1, we can integrate the decision
variables and the parameters under the same season together Q∗i max � 􏽘 δ4τ · Qi max , i ∈ R. (10b)
τ�0
by δ. To make the above model concise, the so-called ‘in-
tegrated’ primal variables are introduced to replace the Other integrated primal variables can be given similarly,
corresponding discounted values. For example, and all of them take an asterisk at their top right corner.
(T/4)− 1 Thus, with the introduction of the ‘integrated’ primal var-
1
QL∗i (j) � 􏽘 δ4τ · QLi (4τ + j), j � 1, 2, 3, 4; i ∈ R, iables, the symbol 􏽐(T/4)−
τ�0 (·) in Objective (9) could be re-
τ�0 moved, and then, the operational stages 5∼T in the submodel
(10a) could be compressed into the stages 1∼4, respectively. Thus,
we have

4
min 􏽘 􏼨 􏽘 Δt · 􏼔ci · xi · 􏼒QL∗i (j) − Q′∗ ∗
i min (j)􏼓 − f · Ki · QEi (j) · Hi 􏼕
j�1 i∈R

+ β+1 (j) · QE∗i (j) − MM · xi 􏼁 + β+2 (j) · Ki · QE∗i (j) · Hi − N∗i max 􏼁 + β−2 (j) · Ki · QE∗i (j) · Hi − xi · N∗i min 􏼁

+ β+3 (j) · y∗i (j) − Cap∗i 􏼁 + β−3 (j) · y∗i (j) − V∗i min · xi 􏼁
(11)
+ β4 (j) · y∗i (j)− y∗i (j − 1)− Δt · Q∗i− 1 (j) + ε∗i (j) − Q∗i (j) − QL∗i (j)􏼁􏼁 + β5 (j) · Q∗i (j) − QE∗i (j) − QA∗i (j)􏼁

+ β+6 (j) · QE∗i (j) + QA∗i (j) − QF∗i max 􏼁 + β−6 (j) · QE∗i (j) + QA∗i (j) − QF∗i min 􏼁

+ β+7 (j) · QL∗i (j) − xi · Q∗i max − 1 − xi 􏼁 · Q∗i min (j)􏼁 + β−7 (j) · QL∗i (j) − Q∗i min (j)􏼁􏼩.

Step 4. Notice that the above transformation is for the Here, we need to extract the relaxed constraints back at first
submodel which is under a specific scenario. Specifically, and then combine the model with each scenario 􏽢s (∈􏽢S) (see
Step 2 relaxes the constraints to get a unconstrainted model, Figure 5) and the planning stage. Thus, we can get the
and then, Step 3 aggregates the multistage for this model. following intermediate model:
Complexity 11

4
min ⎝Δt · 􏽘 r∗ (j) · Cap ⎞
􏽘 e · Capi + 􏽘⎛ ⎠
i i
i∈R j�1 i∈R
(12a)
4
+ 􏽘 p􏽢s · 􏽘 􏽘 Δt · 􏼔ci · xi · 􏼒QL∗i 􏽢s (j) − Q′∗􏽢s ∗􏽢
s
i min (j)􏼓 − f · Ki · QEi (j) · Hi 􏼕,
􏽢s∈􏽢S j�1 i∈R

s.t.
Rmin ≤ 􏽘 xi ≤ R, (12b)
i∈R

Capi ≤ MM · xi , i ∈ R, (12c)

QE∗i 􏽢s (j) ≤ MM · xi , i ∈ R; 􏽢s ∈ 􏽢S; j � 1, 2, 3, 4, (12d)

xi · N∗i min ≤ Ki · QE∗i 􏽢s (j) · Hi ≤ N∗i max , i ∈ R; 􏽢s ∈ 􏽢S; j � 1, 2, 3, 4, (12e)

V∗i min · xi ≤ y∗i 􏽢s (j) ≤ Cap∗i , i ∈ R; 􏽢s ∈ 􏽢S; j � 1, 2, 3, 4, (12f)

y∗i 􏽢s (j) − y∗i 􏽢s (j − 1) � Δt · 􏼔Q∗i−􏽢s1 (j) + ε∗i 􏽢s (j) − Q∗i 􏽢s (j) − QL∗i 􏽢s (j)􏼕,
(12g)
i ∈ R, 􏽢s ∈ 􏽢S; j � 1, 2, 3, 4,

Q∗i 􏽢s (j) � QE∗i 􏽢s (j) + QA∗i 􏽢s (j), i ∈ R; 􏽢s ∈ 􏽢S; j � 1, 2, 3, 4, (12h)

QF∗i min ≤ QE∗i 􏽢s (j) + QA∗i 􏽢s (j) ≤ QF∗i max , i ∈ R; 􏽢s ∈ 􏽢S; j � 1, 2, 3, 4, (12i)

Q∗i 􏽢min
s
(j) ≤ QL∗i 􏽢s (j) ≤ xi · Q∗i max + 1 − xi 􏼁 · Q∗i 􏽢min
s
(j), i ∈ R; 􏽢s ∈ 􏽢S; j � 1, 2, 3, 4, (12j)

xi ∈ {0, 1}, Capi ≥ 0, i ∈ R, (12k)

in which the primal variables Cap∗i can be given as follows: Moreover, it is easy to know that Cap∗i , y∗i 􏽢s (j), QL∗i 􏽢s (j),
(T/4)− 1 Q∗i 􏽢s (j), an dQA∗i 􏽢s (j) are all nonnegative.
QE∗i 􏽢s (j),
Cap∗i � 􏽘 δ4τ · Capi � Capi · 􏼐δ0 + δ4 + δ8 + · · · + δT− 8 + δT− 4 􏼑 Till now, along with the stage reduction as shown in
τ�0 Section 3.1.1, the T-stage operations of the MSSP model is
integrated into a four-stage model. The corresponding
1 − δT 1
� Capi · 4 ≈ Capi · ; i ∈ R, scenario tree of this intermediate model is shown in
1− δ 1 − δ4 Figure 5.
(12l)
in which δT can be approximated as zero when the number of
stages T is large enough. Similarly, the parameters r∗i (j)
could be given as 3.2.2. Step 2 of Dimensionality Reduction. Further model
(T/4)− 1 transformation is required due to the stage reduction in
r∗i (j) � 􏽘 δ4τ · ri (4τ + j), j � 1, 2, 3, 4; i ∈ R. Section 3.1.2. Notice that j (�1, 2, 3, 4) in the intermediate
τ�0 model (12) refers to the first four operational stages. Here, we
(12m) use symbol t to replace the corresponding j and then in-
tegrate stage t � 2, 3, 4 into the first stage (t � 1) as illustrated
We can assume the unit operational cost of candidate i at in Section 3.1.2. The way how to transform is similar to that
each stage t is the same, i.e., ri. Then, (12m) can be further in Section 3.2.1.
represented as We still focus on the operational stage under a specific
1 − δT 1 stagewise scenario. Thus, the corresponding submodel can
r∗i (j) � ri · ≈ ri · , i ∈ R. (12n)
1 − δ4 1 − δ4 be described as follows:
12 Complexity

4
min 􏽘 􏽘􏼚Δt · 􏼔ci · xi · 􏼒QL∗i (t) − Q′∗ ∗
i min (t)􏼓 − f · Ki · QEi (t) · Hi 􏼕􏼛, (13a)
t�1 i∈R

s.t.
(13b)
QE∗i (t) ≤ MM · xi , i ∈ R; t � 1, 2, 3, 4,

xi · N∗i min ≤ Ki · QE∗i (t) · Hi ≤ N∗i max , i ∈ R; t � 1, 2, 3, 4, (13c)

1
V∗i min · xi ≤ y∗i (t) ≤ · Capi , i ∈ R; t � 1, 2, 3, 4, (13d)
1 − δ4

y∗i (t) − y∗i (t − 1) � Δt · 􏼂Q∗i− 1 (t) + ε∗i (t) − Q∗i (t) − QL∗i (t)􏼃,
(13e)
i ∈ R, t � 1, 2, 3, 4,

Q∗i (t) � QE∗i (t) + QA∗i (t), i ∈ R; t � 1, 2, 3, 4, (13f)

QF∗i min ≤ QE∗i (t) + QA∗i (t) ≤ QF∗i max , i ∈ R; t � 1, 2, 3, 4, (13g)

Q∗i min (t) ≤ QL∗i (t) ≤ xi · Q∗i max + 1 − xi 􏼁 · Q∗i min (t), i ∈ R; t � 1, 2, 3, 4, (13h)

Step 5. The discount factor δ is still used to aggregate the

value of the objective function at stage t (�2, 3, 4). Then, the
objective function (13a) is transformed as

4 ⎨
⎧ ⎬
⎫
min 􏽘⎩ δt− 1 · 􏽘 Δt · 􏼂ci · xi · QLi (t) − Q′i min(t)􏼁 − f · Ki · QEi (t) · Hi 􏼃⎭ . (14)
t�1 i∈R

Step 6. Similar to Step 2, use dual multipliers β7+k,t (k � 1, stage t � 1∼4 ((13b)∼(13h), (14)) can be transformed to an
. . ., 7), to relax the k-th one of Constraints (13b)∼(13h) of unconstrainted one:
stage t. Thus, the operational optimization problem under

4
min 􏽘 􏼨δt− 1 · 􏽘 Δt · 􏼔ci · xi · 􏼒QL∗i (t) − Q′∗ ∗
i min (t)􏼓 − f · Ki · QEi (t) · Hi 􏼕
t�1 i∈R

+ β+8,t · QE∗i (t) − MM · xi 􏼁

+ β+9,t · Ki · QE∗i (t) · Hi − N∗i max 􏼁 + β−9,t · Ki · QE∗i (t) · Hi − xi · N∗i min 􏼁
1
+ β+10,t · 􏼒y∗i (t) − · Capi 􏼓 + β−10,t · y∗i (t) − V∗i min · xi 􏼁
1 − δ4
(15)
+ β11,t · y∗i (t) − y∗i (t − 1) − Δt · Q∗i− 1 (t) + ε∗i (t) − Q∗i (t) − QL∗i (t)􏼁􏼁
+ β12,t · Q∗i (t) − QE∗i (t) − QA∗i (t)􏼁
+ β+13,t · QE∗i (t) + QA∗i (t) − QF∗i max 􏼁
+ β−13,t · QE∗i (t) + QA∗i (t) − QF∗i min 􏼁

+ β+14,t · QL∗i (t) − xi · Q∗i max − 1 − xi 􏼁 · Q∗i min (t)􏼁 + β−14,t · QL∗i (t) − Q∗i min (t)􏼁􏼩,
Complexity 13

in which the dual multipliers follow the form

β7+k,t � δt− 1·β7+k. Hence, problem (15) can further be
expressed as

4
min 􏽘 δt− 1 · 􏼨 􏽘 Δt · 􏼔ci · xi · 􏼒QL∗i (t) − Q′∗ ∗
i min (t)􏼓 − f · Ki · QEi (t) · Hi 􏼕
t�1 i∈R

+ β+8 · QE∗i (t) − MM · xi 􏼁

+ β+9 · Ki · QE∗i (t) · Hi − N∗i max 􏼁 + β−9 · Ki · QE∗i (t) · Hi − xi · N∗i min 􏼁

1
+ β+10 · 􏼒y∗i (t) − · Capi 􏼓 + β−10 · y∗i (t) − V∗i min · xi 􏼁
1 − δ4 (16)
+ β11 · y∗i (t) − y∗i (t − 1) − Δt · Q∗i− 1 (t) + ε∗i (t) − Q∗i (t) − QL∗i (t)􏼁􏼁

+ β12 · Q∗i (t) − QE∗i (t) − QA∗i (t)􏼁 + β+13 · QE∗i (t) + QA∗i (t) − QF∗i max 􏼁

+ β−13 · QE∗i (t) + QA∗i (t) − QF∗i min 􏼁

+ β+14 · QL∗i (t) − xi · Q∗i max − 1 − xi 􏼁 · Q∗i min (t)􏼁 + β−14 · QL∗i (t) − Q∗i min (t)􏼁􏼩.

Step 7. Similar to Step 3, the “integrated” primal variables

are introduced to replace the corresponding discounted
values, respectively. Taking QLi(t) and Qimax as examples, we 4
have: 4 Q∗∗
i max � 􏽘 δ
t− 1
· Q∗i max , i ∈ R. (17b)
QL∗∗
i � 􏽘δ
t− 1
· QL∗i (t), i ∈ R, (17a) t�1

t�1
Here, all the primal variables have double asterisks in
their top right corner. Thus,

min 􏽘 􏼨Δt · 􏼔ci · xi · 􏼒QL∗∗ ′∗∗ ∗∗ + ∗∗

i − Qi min 􏼓 − f · Ki · QEi · Hi 􏼕 + β8 · QEi − MM · xi 􏼁
i∈R

+ β+9 · Ki · QE∗∗ ∗∗ − ∗∗ ∗∗
i · Hi − Ni max 􏼁 + β9 · Ki · QEi · Hi − xi · Ni min 􏼁

1
+ β+10 · 􏼒y∗∗
i − · Cap∗i 􏼓 + β−10 · y∗∗ ∗∗ ∗∗ ∗∗ ∗∗ ∗∗ ∗∗ ∗∗
i − Vi min · xi 􏼁 + β11 · yi − yi (0)− Δt · Qi− 1 + εi − Qi − QLi 􏼁􏼁
(18)
1 − δ4

+ β12 · Q∗∗ ∗∗ ∗∗ + ∗∗ ∗∗ ∗∗ − ∗∗ ∗∗ ∗∗
i − QEi − QAi 􏼁 + β13 · QEi + QAi − QFi max 􏼁 + β13 · QEi + QAi − QFi min 􏼁

+ β+14 · QL∗∗ ∗∗ ∗∗ − ∗∗ ∗∗
i − xi · Qi max − 1 − xi 􏼁 · Qi min 􏼁 + β14 · QLi − Qi min 􏼁􏼩,

in which
4 4 4 (T/4)− 1
1 − δ4 y∗∗ t− 1
· y∗i (t) � 􏽘 δt− 1 · 􏽘 δ4τ · yi (4τ + t).
Cap∗i � 􏽘δ t− 1
· Capi � · Capi , i ∈ R, (19) i � 􏽘δ
t�1 1− δ t�1 t�1 τ�0
(20)
and for each i (∈R), y∗∗
i is a discounted value that can be
expressed as From (20), we have
14 Complexity

4 (T/4)− 1 From (21) and (22), we have

δ · y∗∗ t 4τ
i � 􏽘 δ · 􏽘 δ · yi (4τ + t). (21)
t�1 τ�0
y∗∗ ∗∗ 0 T
i (0) � δ · yi + δ · yi (0) − δ · yi (T). (23)

Refer to (20), the discounted value y∗∗ Using (23) to replace y∗∗
i (0) in (18), we have the dis-
i (0) can be
denoted as counted objective as follows:
4 4 (T/4)− 1
y∗∗
i (0) � 􏽘 δ
t− 1
· y∗i (t − 1) � 􏽘 δt− 1 · 􏽘 δ4τ · yi (4τ + t − 1).
t�1 t�1 τ�0

(22)

min 􏽘 􏼨Δt · 􏽨ci · xi · 􏼐QL∗∗ ’∗∗ ∗∗

i − Qi min 􏼑 − f · Ki · QEi · Hi 􏽩
i∈R

+ β+8 · QE∗∗
i − MM · xi 􏼁

+ β+9 · Ki · QE∗∗ ∗∗ − ∗∗ ∗∗
i · Hi − Ni max 􏼁 + β9 · Ki · QEi · Hi − xi · Ni min 􏼁

1 1 − δ4
+ β+10 · 􏼠y∗∗
i − 4 · · Capi 􏼡 + β−10 · y∗∗ ∗∗
i − Vi min · xi 􏼁
1− δ 1− δ (24)
+ β11 · 􏼐y∗∗ ∗∗ T ∗∗ ∗∗ ∗∗ ∗∗
i − δ · yi − yi (0) + δ ·yi (T)− Δt · Qi− 1 + εi − Qi − QLi 􏼁􏼁

+ β12 · Q∗∗ ∗∗ ∗∗
i − QEi − QAi 􏼁

+ β+13 · QE∗∗ ∗∗ ∗∗ − ∗∗ ∗∗ ∗∗
i + QAi − QFi max 􏼁 + β13 · QEi + QAi − QFi min 􏼁

+ β+14 · QL∗∗ ∗∗ ∗∗ − ∗∗ ∗∗
i − xi · Qi max − 1 − xi 􏼁 · Qi min 􏼁 + β14 · QLi − Qi min 􏼁􏼩.

QF∗∗ ∗∗ ∗∗ ∗∗
i min ≤ QEi + QAi ≤ QFi max , i ∈ R, (25g)
Similarly, δT in (24) can be approximated as zero due to
the large T. Then, by extracting the relaxed constraints from
Q∗∗ ∗∗ ∗∗ ∗∗
i min ≤ QLi ≤ xi · Qi max + 1 − xi 􏼁 · Qi min , i ∈ R.
the above objective function, we have
(25h)
∗∗
min 􏽘 Δt · 􏽨ci · xi · 􏼐QLi − Q’∗∗
i min 􏼑 − f · Ki · QE∗∗
i · Hi 􏽩,
i∈R
(25a) Step 8. Combine the above model with each scenario 􏽢s (∈􏽢S),
which is given in Section 3.1.2 and then integrate with the
s.t. planning stage of the original MSSP model. Thus, the final
DE model, which is a two-stage stochastic program, is de-
QE∗∗
i ≤ MM · xi , i ∈ R, veloped as (26a)∼(26k).
(25b) min 􏽘 e · Capi + Δt · 􏽘 r∗∗
i · Capi 􏼁
i∈R i∈R
xi · N∗∗ ∗∗ ∗∗
i min ≤ Ki · QEi · Hi ≤ Ni max , i ∈ R, (25c)
⎨
⎧ ⎬
⎫
+ 􏽘⎩ p􏽢s · 􏽘 Δt · 􏼔ci · xi · 􏼒QL􏽢si − Q′∗∗􏽢s 􏽢s
i min 􏼓 − f · Ki · QEi · Hi 􏼕⎭ ,
1
V∗∗ ∗∗
i min · xi ≤ yi ≤ · Capi , i ∈ R, (25d) 􏽢s∈􏽢S i∈R
1− δ
(26a)
(1 − δ)y∗∗
i − yi (0) � Δt · Q∗∗
i− 1 + ε∗∗
i − Q∗∗
i − QL∗∗
i 􏼁,
s.t.
i ∈ R, (26b)
Rmin ≤ 􏽘 xi ≤ R,
(25e) i∈R

Q∗∗
i � QE∗∗
i + QA∗∗
i , i∈R (25f) Capi ≤ MM · xi , i ∈ R, (26c)
Complexity 15

Tributary 2

The basin in upstream Tributary 3

of Yangtze River

Tributary 1 Candidate 1
Candidate 3
Candidate 2
Demand area
Candidate 4
Tributary 4
Candidate 5

Figure 6: Illustration of hydropower station candidate in upstream of Yangtze river.

Table 3: Related parameters of hydropower station candidates.

Candidate i Ki (kg/(m2·s2)) Nimin (KW) Nimax (KW) Hi (m) ci (RMB/m3) ri (RMB/(m3·year)) Qimax (m3 × 108/h)
1 8.4 2546250 5092500 105 0.1 0.06 0.4676
2 8.4 1506750 3013500 93 0.125 0.072 0.8550
3 8.4 1500000 3000000 90 0.15 0.05 0.5177
4 8.4 3742500 7485000 110 0.175 0.08 0.4662
5 8.4 2625000 5250000 112 0.2 0.1 0.3301

Table 4: Minimum storage and limitation of water flows of hy- Table 6: Predicted water inflows in each seasonal stage.
dropower station candidates.
Candidate Stage (t � 1, 2, 3, 4)
Candidate Vimin QFimin (m3 × 108/ QFimax (m3 × 108/ i Spring Summer Autumn Winter
i (m3 × 108) h) h)
1 1.1059 5.3292 1.9722 0.1678
1 0.0311 0.0648 1.5732 Water inflows 2 0.5126 1.5203 0.5997 0.0817
2 0.0213 0.0468 1.7928 εi (t) (m3 × 107/ 3 0.4926 1.2202 0.6089 0.0813
3 0.0131 0.0432 0.4248 h) 4 0.3003 0.5433 0.2912 0.0725
4 0.0420 0.1620 3.5568 5 0.6649 1.1431 0.5554 0.1446
5 0.0115 0.0396 0.6912

QE􏽢si ≤ MM · xi , 􏽢s ∈ 􏽢S; i ∈ R, (26d)

Table 5: Random water inflows in each seasonal stage of each
hydropower station candidate. 􏽢s
xi · N∗∗ ∗∗
i min ≤ Ki · QEi · Hi ≤ Ni max , 􏽢s ∈ 􏽢S; i ∈ R,
Water inflows (m3 × 107/h)
Candidate i Inflows level (26e)
Spring Summer Autumn Winter
H 1.2846 5.7276 2.2020 0.2031 1
1
V∗∗ 􏽢s 􏽢s ∈ 􏽢S; i ∈ R, (26f)
L 0.9867 4.3995 1.6914 0.1560 i min · xi ≤ yi ≤ · Capi ,
H 0.5932 1.6312 0.6677 0.0985 1− δ
2
L 0.4589 1.2617 0.5165 0.0762
H 0.5653 1.2849 0.6739 0.0969 (1 − δ)y􏽢si − y′i � Δt · 􏼔Q􏽢si− 1 + εi∗∗􏽢s − Q􏽢si − QL􏽢si 􏼕 , 􏽢s ∈ 􏽢S; i ∈ R
3
L 0.4441 1.0093 0.5294 0.0762
H 0.3645 0.5958 0.3356 0.0930 (26g)
4
L 0.2575 0.4209 0.2371 0.0657
5
H 0.6900 1.1642 0.5710 0.1515 Q􏽢si � QE􏽢si + QA􏽢si , 􏽢s ∈ 􏽢S; i ∈ R, (26h)
L 0.6482 1.0936 0.5363 0.1423
H 0.4 0.7 0.55 0.25
Probability 􏽢s 􏽢s
L 0.6 0.3 0.45 0.75 QF∗∗ ∗∗
i min ≤ QEi + QAi ≤ QFi max , 􏽢s ∈ 􏽢S; i ∈ R, (26i)
16 Complexity

Table 7: Predicted water demands in each seasonal stage. After transformation, the planning decisions xi and Capi
Stage (t � 1, 2, 3, 4)
remain unchanged, while the operational stages are aggre-
Candidate
gated. Therefore, the proposed approach not only keeps the
i Spring Summer Autumn Winter
future impacts but also reduces the solving complexity.
1 0.1871 0.2299 0.2196 0.1218
Water
2 0.3420 0.4204 0.4016 0.2228 4. Experimental Study
demands
3 0.2071 0.2546 0.2432 0.1349
Qi min (t)
4 0.1865 0.2292 0.2189 0.1215 To validate the applicability of our proposed approach, we
(m3 × 106/h)
5 0.1321 0.1623 0.1551 0.0860 give the data of a case in Section 4.1 at first. Then, based on
this case, we compare our approach with the MSSP model in
the small-scale setting in Section 4.2. Furthermore, we apply
Table 8: Estimation gaps of the DE model. our approach to an overlong-term case in Section 4.3 and
Planning stages finally implement sensitivity analysis in Section 4.4.
Discount factor δ
T�4 T�8 T �12
0.99 25.53% 22.51% 19.94% 4.1. Description of the Case. Yangtze River flows through 11
0.95 17.89% 6.71% − 2.99% provinces in China with a drainage area of 1.8 million square
0.90 8.72% − 9.76% − 24.32% kilometers. The whole river can be divided into the up-
0.85 − 0.51% − 23.20% − 39.73% stream, midstream, and downstream, among which the
0.80 − 9.16% − 34.34% − 50.97% basin with a length of 1030 km in upstream is studied here.
0.75 − 17.29% − 43.92% − 59.59%
According to Wang et al. [45], five hydropower station
0.70 − 24.72% − 51.99% − 66.37%
0.65 − 31.44% − 58.62% − 71.61% candidates could be considered in this basin. Its schema-
0.60 − 37.53% − 64.02% − 75.64% tization is shown in Figure 6 which also involves the main
tributaries.
Among the five hydropower station candidates, at least
Table 9: Average computational time (ms) of both models. two of them should be chosen in the plan. Unit construction
fee (e) and unit profit of power generation (f ) in Table 1 are
Model T�4 T�8 T �12
set to be 5.09 (RMB/m3) and 0.25 (RMB/KWh), respectively
DE model 1224.00 1206.44 1170.22 [46]. Related parameters regarding five candidates are dis-
MSSP model 300.22 3095.56 159373.22
played in Table 3. Once the hydropower station candidate is
chosen, the planned lowest water storage and upper and
lower limit of water discharge are shown in Table 4. The
Table 10: Probabilistic distributions of water inflows and demands initial water flows from upstream to each hydropower
in each seasonal stage. station candidate is set to be 0.2141, 0.0988, 0.0942, 0.0429,
Water inflows (p1H, p1M, Water demands (p2H, and 0.1080 (m3 × 108/h), respectively. And the initial water
Stage storage of each chosen hydropower station is assumed as the
p1L) p2L)
Spring (0.2, 0.4, 0.4) (0.5, 0.5)
twice of the lowest water storage.
Summer (0.5, 0.4, 0.1) (0.65, 0.35)
Autumn (0.3, 0.5, 0.2) (0.6, 0.4) 4.2. Comparison with the MSSP Model by Small-Scale Cases.
Winter (0.1, 0.3, 0.6) (0.3, 0.7) The developed DE model is an approximate version of the
MSSP model (1). In this section, we aim to evaluate the
performance of our approach by comparing the planning
Qi∗∗􏽢s 􏽢s ∗∗ ∗∗􏽢s
􏽢s ∈ 􏽢S; i ∈ R, decisions as well as the operational performance with these
min ≤ QLi ≤ xi · Qi max + 1 − xi 􏼁 · Qi min ,
of the MSSP model. Due to the curse of dimensionality faced
(26j) by the MSSP model, we consider small-scale cases based on
Section 4.1.
Capi , y􏽢si , QL􏽢si , QE􏽢si , Q􏽢si , QA􏽢si ≥ 0, xi ∈ {0, 1}, i ∈ R; 􏽢s ∈ 􏽢S,
(26k) 4.2.1. Parameters in Small-Scale Experiments. To limit the
in which the discounted value of water inflows and demands dimensionality here, we set the planning horizon as 1, 2, and
∗∗􏽢 3 years, respectively. According to the seasonal features of
εi∗∗􏽢s and Qimin
s
(i ∈ R) can be obtained similar to the way used
in formula (3). Besides, r∗∗ which, one year is divided into four stages: spring (March to
i in (26a) is an integrated variable:
May), summer (June to August), autumn (September to
4
r∗∗ t− 1
· r∗i (t), (26l) November), and winter (December to February). Then, each
i � 􏽘δ i ∈ R.
t�1
stage has 2160 (h) (i.e., Δt). Thus, the number of operational
stages T is 4, 8, and 12, respectively. Besides, we merely
Moreover, for simplicity, the double asterisks of the consider the uncertainty caused by water inflows. Two kinds
decision variables are omitted. of stochastic outside water inflows (H, L), as well as the
Complexity 17

3.5
High Middle Low

Precipitation (m3 × 107/h)

3

2.5

1.5

0.5

0
Spring Summer Autumn Winter Spring Summer Autumn Winter Spring Summer Autumn Winter
1 0.8675 3.8679 1.487 0.1372 0.5451 2.4303 0.9343 0.0862 0.2226 0.9927 0.3816 0.0352
2 0.5468 1.5035 0.6155 0.0908 0.3702 1.0181 0.4168 0.0615 0.1937 0.5327 0.2181 0.0322
3 0.5412 1.2302 0.6452 0.0928 0.3724 0.8464 0.4439 0.0639 0.2035 0.4625 0.2426 0.0349
4 0.4292 0.7015 0.3951 0.1095 0.3968 0.6487 0.3654 0.1013 0.3645 0.5958 0.3356 0.093
5 0.3378 0.57 0.2795 0.0742 0.2912 0.4913 0.2409 0.0639 0.2445 0.4126 0.2023 0.0537
Stage

Figure 7: High, middle, low-level precipitation in each seasonal stage of each hydropower station candidate.

5
Middle
Tributary (m3 × 107/h)

High Low
4

0
Spring Summer Autumn Winter Spring Summer Autumn Winter Spring Summer Autumn Winter
1 1.2735 5.6781 2.183 0.2014 1.0188 4.5425 1.7464 0.1611 0.7641 3.4068 1.3098 0.1208
2 0.4419 1.2151 0.4974 0.0734 0.3535 0.9721 0.3979 0.0587 0.2651 0.7291 0.2984 0.044
3 0.401 0.9113 0.478 0.0688 0.3208 0.7291 0.3824 0.055 0.2406 0.5468 0.2868 0.0413
4 0 0 0 0 0 0 0 0 0 0 0 0
5 0.7425 1.2527 0.6144 0.163 0.594 1.0022 0.4915 0.1304 0.4455 0.7516 0.3686 0.0978
Stage

Figure 8: High, middle, low-level tributary in each seasonal stage of each hydropower station candidate.

corresponding probabilities (p1H, p1L), in each seasonal stage 4.2.2. Computational Results. We optimally solve the DE
are given in Table 5. The constant water inflows and water model and the MSSP model by IBM ILOG CPLEX 12.6.3 on
demands used in the simplification are displayed in Tables 6 a PC with 8 GB memory and a CPU at 2.5 GHz.
and 7, respectively. The acceptance level of the water spill is It is found that the DE model can generate the same
assumed as twice of the demands. planning results with the MSSP model for all these small-
Because the random inflows fluctuate with seasons, we scale cases. Besides, in terms of the operations, the DE
assume the water inflows of each candidate take the same model considers the multistage operational impacts by the
uncertain level in each season. Thus, for the MSSP model, aggregated way. Hence, it is interesting to observe how
16(�24), 256(�28), and 4096(�212) stagewise scenarios are accurate the DE model is in the estimation of the future
constructed, respectively. Furthermore, according to the operational performance of the cascade hydropower
generation and simplification process of the scenario set system. We use DE_2nd and MSSP_2nd to denote the
displayed in Section 3.1, 16 scenarios are generated for the operational performance generated by the DE model and
two-stage DE model. Assume that spring is the first oper- the MSSP model, respectively. Table 8 displays the esti-
ational season. The discount factor δ used in the DE model is mation gap (�(DE_2nd-MSSP_2nd)/MSSP_2nd) under
set to be 0.99, 0.95, . . ., 0.60, respectively. different discount factors.
18 Complexity

0.6

0.5
High Low

Demand (m3 ×106/h)

0.4

0.3

0.2

0.1

0
Spring Summer Autumn Winter Spring Summer Autumn Winter
1 0.2368 0.2703 0.2644 0.1722 0.1433 0.1634 0.1599 0.1045
2 0.4275 0.4888 0.4781 0.3094 0.2565 0.2933 0.2868 0.1856
3 0.2589 0.296 0.2895 0.1873 0.1553 0.1776 0.1737 0.1124
4 0.2331 0.2665 0.2606 0.1687 0.1399 0.1599 0.1564 0.1012
5 0.1651 0.1887 0.1846 0.1195 0.099 0.1132 0.1107 0.0717
Stage

Figure 9: High, low-level water demand in each seasonal stage of each hydropower station candidate.

Table 11: Predicted water inflows and demands in each seasonal stage.
Stage (t � 1, 2, 3, 4)
Candidate i
Spring Summer Autumn Winter
1 1.4484 8.0021 2.7796 0.2017
2 0.6708 2.2816 0.8445 0.0982
Water inflows εi (t) (m3 × 107/h) 3 0.6433 1.8019 0.8559 0.0975
4 0.3904 0.6698 0.3683 0.0971
5 0.8462 1.6252 0.7486 0.1729
1 0.1871 0.2299 0.2196 0.1218
2 0.3420 0.4204 0.4016 0.2228
Water demands Qi min (t) (m3 × 106/h) 3 0.2071 0.2546 0.2432 0.1349
4 0.1865 0.2292 0.2189 0.1215
5 0.1321 0.1623 0.1551 0.0860

Table 12: Planning decisions and the cost/benefit.

3 8
Candidate i Capi (m × 10 ) Operational cost (RMB × 108) Penalty cost (RMB × 108) Benefit (RMB × 108)
1 0.214 0.063 347.191 550.002
4 0.043 0.017 0 808.401
5 0.108 0.053 0 567.013

Table 13: Operational decisions under scenario 16.

Candidate i 1 2 3 4 5
3 8
QL16i (m × 10 /h) 2.154 0.066 0.040 0.036 0.026
16 3
QEi (m /h) 115476 0 0 162013 111607
3 8
QA16 i (m × 10 /h) 4.690 6.714 8.496 9.246 11.014
3 8
y16
i (m × 10 ) 4.282 0 0 0.858 2.161
3 8
Q16
i (m × 10 /h) 4.691 6.714 8.496 9.248 11.015

Table 14: Operational decisions under scenario 1281.

Candidate i 1 2 3 4 5
QL1281
i (m3 × 108/h) 0.874 0.073 0.044 0.040 0.028
QE1281
i (m3/h) 115476 0 0 162013 111607
QAi 1281
(m3 × 108/h) 5.171 6.936 8.496 9.219 10.847
3
y1281
i (m × 108) 4.282 0 0 0.858 0.230
Qi1281
(m × 108/h)
3
5.172 6.936 8.496 9.221 10.848
Complexity 19

Table 15: Corresponding cost terms and benefit.

Operational Candidate Total performance Operational cost Penalty cost Power generation
stages i (RMB × 108) (RMB × 108) (RMB × 108) (RMB × 108)
1 471.396 0.063 77.453 550.002
4 808.165 0.017 0 808.401
T � 40
5 566.410 0.053 0 567.013
Total 1845.971 0.134 77.453 1925.416
1 236.358 0.063 312.491 550.002
4 808.165 0.017 0 808.401
T � 80
5 566.410 0.053 0 567.013
Total 1610.933 0.134 312.491 1925.416
1 205.604 0.063 343.245 550.002
4 808.165 0.017 0 808.401
T �120
5 566.410 0.053 0 567.013
Total 1580.179 0.134 343.245 1925.416
1 201.653 0.063 347.196 550.002
4 808.165 0.017 0 808.401
T �160
5 566.410 0.053 0 567.013
Total 1576.228 0.134 347.196 1925.416
1 201.150 0.063 347.699 550.002
4 808.165 0.017 0 808.401
T � 200
5 566.410 0.053 0 567.013
Total 1575.725 0.134 347.699 1925.416

1e7 1e12 1e9

60
Total performance/profit of power

7.4 0.9
generation/penalty (RMB)

0.8 50
Operational cost (RMB)

6.4
0.7
Penalty cost (RMB)

5.4 40
0.6
4.4 0.5
0.4 30
3.4
0.3
2.4 20
0.2
1.4 0.1
10
0.4 0
δ = 0.99

δ = 0.98

δ = 0.97

δ = 0.96

δ = 0.95

δ = 0.94

δ = 0.93

δ = 0.92

δ = 0.91

δ = 0.90

δ = 0.89

δ = 0.88

0
40 80 120 160
Change of periods
Change of the discount factor
δ = 0.99 δ = 0.96 δ = 0.93 δ = 0.89
Operational cost Profit of power generation
δ = 0.98 δ = 0.95 δ = 0.92 δ = 0.90
Total performance Penalty cost δ = 0.97 δ = 0.94 δ = 0.91 δ = 0.88
Figure 10: Cost and profit of power generation under the different Figure 11: Penalty cost under the different discount factors and
discount factors. planning horizons.

of the future operational performance by choosing a suitable

It is found that the operational performance gener- discount factor, while less calculational time is required.
ated by the DE model significantly varies with the dis-
count factors. Specifically, when δ is large, more future
4.3. Real Case Study. Section 4.2 validates the effectiveness of
impacts would be taken into account which results in the
our proposed DE model in the small-scale setting. In this
overestimation. Hence, the estimation gap is positive.
section, we further apply our approach to a large-scale case
With the decrease of δ, less future performance would be
to illustrate its applicability.
discounted. Thus, the estimation gap tends to be negative
gradually. It also means that the DE model can generate
accurate estimation of the future operational perfor- 4.3.1. Parameters in the Large-Scale Case. Based on the case
mance by choosing the appropriate discount factor. given in Section 4.1, we consider the 40-year planning
Moreover, the computational time of the DP model horizon (i.e., T �160). Three kinds of stochastic water in-
outperforms that of the MSSP model greatly especially when flows (H, M, and L) and two kinds of stochastic water de-
the horizon is large, as shown in Table 9. mands (H, L) are taken into account in each seasonal stage.
In summary, the DP model can yield the same location The corresponding probabilities of two random variables are
decisions with the MSSP model and the accurate estimation (p1H, p1M, p1L) and (p2H, p2L) as shown in Table 10. Refer to
20 Complexity

1e9 which represents the low-level inflows and the high-level

910 demands, requires hydropower stations to utilize their
810 storage efficiently and then results in low water storage.
Total performance (RMB)

710 Furthermore, in the computational results, the scheduling

610 strategies would ensure that the values of water discharge
510 (Q) are between the upper and lower limits in both scenarios,
410 making it capable of preventing flood and drainage. Notice
310 that the water discharges of power generation under two
210 scenarios are the same and the upper limit of output has
110 been reached under both scenarios. Thus, the scheduling
10 strategies can always pursue maximum benefits of power
40 80 120 160
Change of periods generation under different scenarios. In such a way, the
proposed DE model can optimize the global performance in
δ = 0.99 δ = 0.96 δ = 0.93 δ = 0.89
the uncertain setting.
δ = 0.98 δ = 0.95 δ = 0.92 δ = 0.90
δ = 0.97 δ = 0.94 δ = 0.91 δ = 0.88

Figure 12: Total performance under the different discount factors 4.4. Sensitivity Analysis. In this section, we carry out the
and planning horizons. sensitivity analysis to reveal how critical parameters influ-
ence the strategic decisions and the performance of the
cascade hydropower stations. Because one of the merits is to
Wang et al. [47] and Yang and Zhang [48]. Uncertain water handle overlong stochasticity by the idea of aggregation, two
inflows which mainly include precipitation and water from natural concerns are how long and how much are dis-
main tributaries (drainage areas larger than ten thousand counted. Here, we focus on the planning horizon (T) and the
square kilometers) are given in Figures 7 (precipitation) and discount factor (δ).
8 (tributary), respectively. Random demands are given in
Figure 9. Besides, the predicted water inflows and demands
used in the DE model are shown in Table 11, and the random 4.4.1. Sensitivity Analysis of the Planning Horizon. The
acceptance level of water spill is also assumed as twice of the planning horizon (T) is one of the main factors causing the
corresponding demands. curse of dimensionality [19, 41]. The DE model aggregates
Thus, the number of stagewise scenarios in the MSSP the multiple operational stages, but T would influence the
model is up to 6160(≈3.193 × 10124), which is impossible to results of the model transformation. Hence, we focus on the
handle exactly. By the DE model, the number of scenarios is length of the planning horizon here. Considering the long-
reduced to 1296. The discount factor δ is set to be 0.95 here. lasting effects of the location decisions, we set the length of
the planning horizon as 10, 20, 30, 40, and 50 years, re-
spectively. That is, we consider T � 40, 80, 120, 160, and 200.
4.3.2. Computational Results of the DE Model. The optimal Other parameters are the same as those in Section 4.3.
solution of the DE model can be obtained by within 2047 After calculation, it is found that the decisions of location
(ms) by IBM ILOG CPLEX 12.6.3, in which Candidates 1, 4, and capacity under different T, which are the planning
and 5 are selected. Table 12 shows the capacity decisions and results we care about in our study, are the same as these of
the operational performance, including the operational cost, T �160 (see Table 12). It means these planning horizons
the penalty cost, and the profit of power generation, of each would not alter the strategic decisions of the DE model.
selected candidate. The total performance value including Furthermore, the cost terms and the profit of power
the planning stage and the operational stages is 1.576 × 1011 generation are displayed in Table 15. It is found that the
(RMB). operational cost and the profit of power generation remain
We further observe two extreme scenarios. One is sce- unchanged while the penalty approaches to be stable with
nario 16 (high-level inflows and low-level demands in all the increase of T. Specifically, the operational cost is the
four seasons), and another scenario is 1281 (low-level in- capacity-related cost within the planning horizon. In the DE
flows and high-level demands in all four seasons). The model, the impacts of the capacity-related cost are aggre-
corresponding scenario-dependent operational results of gated by δT, respectively. When T is extremely large, δT can
five hydropower station candidates under two scenarios are be approximated as zero, see (12l) and (12n). Hence, the
shown in Tables 13 and 14, respectively. operational cost keeps unchanged. Besides, to calculate the
As shown in Tables 13 and 14, the daily scheduling benefit of the power generation, it is found that the maxi-
strategies would be adjusted according to different scenarios. mum output is reached in our cases. Thus, after aggregation
First, in terms of loss flow (QL), when scenario 16 happens, by δT, the benefit of the power generation also remains
the value is relatively high in hydropower station 1 to constant. Moreover, the penalty of hydropower station 4 and
prevent the flood (see Table 13). While if scenario 1281 5 are zero because the loss flows are smaller than the cor-
occurs, the loss flow merely satisfies demand (see Table 14). responding penalty thresholds. However, T would influence
Then, regarding water storage (y), the value is high in hy- y∗∗
i (see (20)) at first and then the loss flow by constraints
dropower station 5 under scenario 16 due to the high-level (25e). Hence, the penalty of hydropower station 1 would
inflows and the low-level demands. However, scenario 1281, increase gradually until coming to be stable.
Complexity 21

Based on the above three terms, with the increase of T, Figure 11 shows the varying pattern of the penalty with
the total performance also shows a trend of increasing first the decrease of δ under each T. Specifically, the penalty
and then approaches to be stable. This result coincides with increases first and then decreases. The reason is the same as
the essence of the DE model. Specifically, the farther the that in Figure 10. However, with the increase of T, the peak
future is, the fewer the impacts would be taken into account. point would appear earlier. It is because bigger T can bring
Thus, the results would close to be stable finally. It is also in significant discounting effects due to δT.
accord with the decision-making in the real-world, i.e., the Furthermore, we show the change of the total perfor-
coming future would be put more emphasis than the mance in Figure 12. The variations under different T are very
superfar future. Hence, when T is large enough, the calcu- similar. Specifically, with the decline of δ, the total perfor-
lational results of the DE model would tend to be stable. It mance gradually decreases and finally tends to be stable.
means that the DE model could describe the steady results of Besides, as δ becomes smaller, the total performance would
long-term stochastic operations. Thus, for the decision- reach its steady state within a shorter planning horizon,
makers (DMs), it is not necessary to increase the planning specifically, around 120 periods when δ � 0.96 and around
horizon infinitely. merely 40 periods when δ � 0.90. Thus, we conclude that
when the low δ is considered, the DMs could only consider a
relatively shorter planning horizon since the impacts in the
4.4.2. Sensitivity Analysis of the Discount Factor.
long-term future would be significantly discounted.
Discount factor (δ) is an exogenous variable introduced by
the DE model to help to reduce dimensionality. In Section
4.2, we have shown that the DE model can generate accurate 5. Conclusions and Future Research Directions
operational performance estimation by selecting appropriate
The integrated planning of cascade hydropower stations,
δ. In this section, we further explore the impacts of the
which takes the operational strategies under future un-
discount factor in the large-scale setting. We vary it as 0.99,
certainty into account, can bring global benefit. An MSSP
0.98, . . ., 0.88. Other parameters are the same as those of
model is developed to deal with the location and capacity
Section 4.3.
decisions of cascade hydropower stations with the consid-
The calculational results show the discount factors in the
eration of the expected long-term performance. To address
value range above will not influence the final strategic de-
the computational issue caused by the overlong stochasticity,
cisions. Besides, the operational cost, the penalty cost, the
a dimensionality reduction approach based on DE is
profit of power generation, and the total performance under
designed to transform the MSSP model into a two-stage
different discount factors are shown in Figure 10. Firstly, it is
stochastic programming model, in which future impacts
shown that, with the decrease of the discount factor, both the
remain by aggregating the operational stages together. Thus,
operational cost and the profit decline with a decreasing
the optimization of the large-scale hydropower stations
speed since they are discounted more. Moreover, in terms of
could be addressed regardless of the number of operational
the penalty cost, it remains zero when δ is large (0.99). This is
stages.
because the discounted penalty threshold also decreases but
Experiments based on Yangtze River in China are
is still larger than the loss flow. Then, the penalty cost in-
implemented. First, by the comparison with the MSSP
creases significantly with the decline of δ to 0.98. The reason
model under the small-sized cases, it is found that our
is that the discounted penalty threshold begins to be smaller
approach can obtain the same planning results and accurate
than the loss flow, leading to the larger penalty. Finally, with
estimation for future operational performance with a short
its continuous decrease (less than 0.98), δ dominates the
computational time. Then, a large-scale case is studied to
penalty term which makes the aggregated penalty smaller.
show how the approach is applied to solve a practical
These results indicate different discount factors will not
problem. Moreover, to evaluate the influence of critical
influence the location decisions but result in the variation of
parameters in the proposed DE model, sensitivity analysis is
the operational performance. Hence, the DMs should
conducted based on the planning horizon and the discount
carefully choose the discount factor if they want to get an
factor. The results reveal that our model is robust to different
accurate estimation of the operational performance.
parameter settings to obtain stable location-related planning
results, while the estimation of the future performance
4.4.3. Mutual Effects of Two Factors. We further investigate would be influenced by the planning horizon and the dis-
the mutual impacts of two factors on decisions and objective count factor. The corresponding observations and impli-
terms. We set T as 40, 80, 120, 160, and 200 and δ as 0.99, cations are also obtained from the results.
0.98, . . ., 0.88 and then calculate the model under each In the future, we can extend our study in the following
combination of two kinds of parameters. aspects. First, the discount factor plays an important role in
From the calculational results, we find that the locations our study. Specifically, it not only discounts the future
and capacities of hydropower stations remain unchanged impacts to get the so-called net present value but also has the
under the different combinations of planning horizon and internal relationship between the dual multipliers. Thus, it
discount factor. Given a δ, the operational cost and the profit influences the accuracy of the future estimation of the DE
of power generation will not be influenced by the change of model. In this paper, we follow the way of Grinold [42] and
T, which is similar to Table 15. Hence, we focus on the King and Wallace [43] to give the discount factor directly.
penalty and the total operational performance. But, how to select the appropriate discount factor to get
22 Complexity

better performance estimation should be considered in the [7] Y. Xu and Y. Mei, “A modified water cycle algorithm for long-
future study. It is a hard issue but deserves further effort. term multi-reservoir optimization,” Applied Soft Computing,
Second, the minimum installed power and the nonlinearity, vol. 71, pp. 317–332, 2018.
like the stage-storage curve, are ignored in this paper. The [8] D. G. Larentis, W. Collischonn, F. Olivera, and C. E. M. Tucci,
consideration of these requirements and constraints could “GIS-based procedures for hydropower potential spotting,”
Energy, vol. 35, no. 10, pp. 4237–4243, 2010.
help to improve the applicability of our approach to practical
[9] B. C. Kusre, D. C. Baruah, P. K. Bordoloi, and S. C. Patra,
problems, which leaves certain spaces for future study. Fi- “Assessment of hydropower potential using GIS and hy-
nally, in this paper, the impacts of time-varying stochastic drological modeling technique in Kopili river basin in Assam
operations are considered in the form of expected value. (India),” Applied Energy, vol. 87, no. 1, pp. 298–309, 2010.
Another popular way is to formulate the objective of Value- [10] B. Serpoush, M. Khanian, and A. Shamsai, “Hydropower plant
at-Risk or Conditional Value-at-Risk to incorporate the site spotting using geographic information system and a
DMs’ risk attitudes. Then, is it possible to utilize the MATLAB based algorithm,” Journal of Cleaner Production,
characteristics of Value-at-Risk or Conditional Value-at- vol. 152, pp. 7–16, 2017.
Risk to facilitate our scenario tree reduction? It is another [11] A. Z. Zaidi and M. Khan, “Identifying high potential locations
interesting research direction. for run-of-the-river hydroelectric power plants using GIS and
digital elevation models,” Renewable and Sustainable Energy
Reviews, vol. 89, pp. 106–116, 2018.
Data Availability [12] J. Hosnar and A. Kovač-Kralj, “Mathematical modelling and
MINLP programming of a hydro system for power genera-
All data included in this paper are available upon request by tion,” Journal of Cleaner Production, vol. 65, no. 4, pp. 194–
contact with the corresponding author. 201, 2014.
[13] C. Loannidou and J. R. O’Hanley, “Eco-friendly location of
Conflicts of Interest small hydropower,” European Journal of Operational Re-
search, vol. 264, pp. 907–918, 2018.
The authors declare that there are no conflicts of interest [14] D. A. G. Vieira, L. S. M. Guedes, A. C. Lisboa, and
regarding the publication of this paper. R. R. Saldanha, “Formulations for hydroelectric energy
production with optimality conditions,” Energy Conversion
and Management, vol. 89, pp. 781–788, 2015.
Acknowledgments [15] C. T. Cheng, J. J. Shen, and X. Y. Wu, “Short-term scheduling
for large-scale cascaded hydropower systems with multi-
The authors wish to express their gratitude to the five
vibration zones of high head,” Journal of Water Resource
anonymous referees for their constructive criticism, which led Planning and Management, vol. 138, no. 3, pp. 257–267, 2012.
to substantial improvements of the paper. This research was [16] C. Li, J. Zhou, S. Ouyang, X. Ding, and L. Chen, “Improved
partly supported by the National Natural Science Foundation decomposition-coordination and discrete differential dy-
of China (Grant no. 71971053, 71832001, 71572033) and the namic programming for optimization of large-scale hydro-
MOE (Ministry of Education in China) Project of Humanities power system,” Energy Conversion and Management, vol. 84,
and Social Sciences (Grant no. 18YJA630129). pp. 363–373, 2014.
[17] X. Li, J. Wei, T. Li, G. Wang, and W. W.-G. Yeh, “A parallel
References dynamic programming algorithm for multi-reservoir system
optimization,” Advances in Water Resources, vol. 67, no. 4,
[1] A. Afshar, F. Sharifi, and M. R. Jalali, “Non-dominated ar- pp. 1–15, 2014b.
chiving multi-colony ant algorithm for multi-objective op- [18] Z.-K. Feng, W.-J. Niu, C.-T. Cheng, and S.-L. Liao, “Hy-
timization: application to multi-purpose reservoir operation,” dropower system operation optimization by discrete differ-
Engineering Optimization, vol. 41, no. 4, pp. 313–325, 2009. ential dynamic programming based on orthogonal
[2] P. Fraenkel, O. Paish, V. Bokalders, A. Harvey, and experiment design,” Energy, vol. 126, pp. 720–732, 2017.
A. E. Brown, Micro-Hydro Power: A Guide for Development [19] Z.-K. Feng, W.-J. Niu, C.-T. Cheng, and X.-Y. Wu, “Opti-
Workers, IT Publications Ltd., London, UK, 1999. mization of hydropower system operation by uniform dy-
[3] G. Eiger and U. Shamir, “Optimal operation of reservoirs by namic programming for dimensionality reduction,” Energy,
stochastic programming,” Engineering Optimization, vol. 17, vol. 134, pp. 718–730, 2017.
no. 4, pp. 293–312, 1991. [20] C. Cheng, J. Shen, X. Wu, and K.-W. Chau, “Short-term
[4] A. Gjelsvik, B. Mo, and A. Haugstad, “Long- and medium- hydroscheduling with discrepant objectives using multi-step
term operations planning and stochastic modelling in hydro- progressive optimality algorithm,” JAWRA Journal of the
dominated power systems based on stochastic dual dynamic American Water Resources Association, vol. 48, no. 3,
programming,” in Handbook of Power Systems I; Energy pp. 464–479, 2012.
Systems; Section 2, Springer, Berlin, Germany, 2010. [21] Z.-K. Feng, W.-J. Niu, and C.-T. Cheng, “Optimizing elec-
[5] C. Cheng, S. Wang, K.-W. Chau, and X. Wu, “Parallel discrete trical power production of hydropower system by uniform
differential dynamic programming for multireservoir oper- progressive optimality algorithm based on two-stage search
ation,” Environmental Modelling & Software, vol. 57, mechanism and uniform design,” Journal of Cleaner Pro-
pp. 152–164, 2014. duction, vol. 190, pp. 432–442, 2018.
[6] C. Wang, J. Zhou, P. Lu, and L. Yuan, “Long-term scheduling [22] W.-J. Niu, Z.-K. Feng, C.-T. Cheng, and X.-Y. Wu, “A parallel
of large cascade hydropower stations in Jinsha river, China,” multi-objective particle swarm optimization for cascade hy-
Energy Conversion and Management, vol. 90, no. 15, dropower reservoir operation in southwest China,” Applied
pp. 476–487, 2015. Soft Computing, vol. 70, pp. 562–575, 2018.
Complexity 23

[23] G. Zhang, Y. Wu, F. Zhang, and X. Liu, “Application of [39] M. N. Hjelmeland, A. Helseth, and M. Korpås, “Medium-term
adaptive quantum particle swarm optimization algorithm for hydropower scheduling with variable head under inflow,
optimal dispatching of cascaded hydropower stations,” in energy and reserve capacity price uncertainty,” Energies,
Intelligent Computing Theories and Applications, Lecture vol. 12, pp. 1–15, 2019.
Notes in Computer Science, Springer, Berlin, Germany, 2012. [40] H. Poorsepahy-Samian, V. Espanmanesh, and B. Zahraie,
[24] J. Zhang, Q. Tang, Y. Chen, and S. Lin, “A hybrid particle “Improved inflow modeling in stochastic dual dynamic
swarm optimization with small population size to solve the programming,” Journal of Water Resources Planning and
optimal short-term hydro-thermal unit commitment prob- Management, vol. 142, no. 12, Article ID 04016065, 2016.
lem,” Energy, vol. 109, pp. 765–780, 2016. [41] Z.-K. Feng, W.-J. Niu, and C.-T. Cheng, “China’s large-scale
[25] S. Ouyang, J. Zhou, C. Li, X. Liao, and H. Wang, “Optimal hydropower system: operation characteristics, modeling
design for flood limit water level of cascade reservoirs,” Water challenge and dimensionality reduction possibilities,” Re-
Resources Management, vol. 29, no. 2, pp. 445–457, 2015. newable Energy, vol. 136, pp. 805–818, 2019.
[26] C.-T. Tseng, Y.-L. Lee, and C.-C. Chou, “An ambient in- [42] R. C. Grinold, “Model building techniques for the correction
telligence system for water release scheduling in cascade of end effects in multistage convex programs,” Operations
hydropower systems,” Journal of Ambient Intelligence and Research, vol. 31, no. 3, pp. 407–431, 1983.
Humanized Computing, vol. 6, no. 1, pp. 47–55, 2015. [43] A. J. King and S. W. Wallace, Modeling with Stochastic
[27] Z. K. Feng, W. J. Niu, J. Z. Zhou, C. T. Cheng, H. Qin, and Programming, Springer International Publishing, New York,
Z. Q. Jiang, “Parallel multi-objective genetic algorithm for USA, 2012.
short-term economic environmental hydrothermal schedul- [44] A. K. Konicz, D. Pisinger, K. M. Rasmussen, and
ing,” Energies, vol. 10, no. 2, pp. 1–22, 2017. M. Steffensen, “A combined stochastic programming and
[28] W. J. Niu, Z. K. Feng, B. F. Feng, Y. W. Min, C. T. Cheng, and optimal control approach to personal finance and pensions,”
J. Z. Zhou, “Comparison of multiple linear regression, arti- OR Spectrum, vol. 37, no. 3, pp. 583–616, 2015.
ficial neural network, extreme learning machine, and support [45] M. L. Wang, X. Y. Wan, P. A. Zhong, Y. Zhang, and C. Liu,
vector machine in deriving operation rule of hydropower “Characteristics of precipitation in upper reaches of the
reservoir,” Water, vol. 11, pp. 1–17, 2019. Yangtze river and its temporal and spatial evolution,” South-
[29] Q. Mao, S. F. Mueller, and H.-M. H. Juang, “Quantitative to-North Water Transfers and Water Science & Technology,
precipitation forecasting for the Tennessee and cumberland vol. 14, no. 4, pp. 65–71, 2016.
river watersheds using the NCEP regional spectral model,” [46] M. Zhang, C. J. Li, Y. C. Zhang, and X. H. Yuan, “Application
Weather and Forecasting, vol. 15, no. 1, pp. 29–45, 2000. of minimum risk control model to hydropower generation
[30] C. E. M. Tucci, R. T. Clarke, W. Collischonn, P. L. D. S. Dias, scheduling,” Journal of Huazhong University of Science and
and G. S. D. Oliveira, “Long-term flow forecasts based on Technology (Natural Science Edition), vol. 36, no. 9, pp. 25–28,
climate and hydrologic modeling: Uruguay river basin,” 2008.
Water Resources Research, vol. 39, no. 7, p. 1181, 2003. [47] Y. Wang, K. L. Xin, P. Zhao, and J. R. Zhou, “Prediction of
[31] B. Xu, P.-A. Zhong, R. C. Zambon, Y. Zhao, and seasonal water consumption by Winters method,” Water and
W. W.-G. Yeh, “Scenario tree reduction in stochastic pro- Wastewater Engineering, vol. 37, no. s1, pp. 183–186, 2011.
gramming with recourse for hydropower operations,” Water [48] Y. D. Yang and Y. Z. Zhang, “On water use situation in the
Resources Research, vol. 51, no. 8, pp. 6359–6380, 2015. Yangtze river basin,” Yangtze River, vol. 37, no. 8, pp. 10-11,
[32] S.-E. Fleten and T. K. Kristoffersen, “Short-term hydropower 2006.
production planning by stochastic programming,” Computers
& Operations Research, vol. 35, no. 8, pp. 2656–2671, 2008.
[33] M. Chazarra, J. Garcı́a-González, J. I. Pérez-Dı́az, and
M. Arteseros, “Stochastic optimization model for the weekly
scheduling of a hydropower system in day-ahead and sec-
ondary regulation reserve markets,” Electric Power Systems
Research, vol. 130, pp. 67–77, 2016.
[34] B. Xu, P.-A. Zhong, Y. Wu, F. Fu, Y. Chen, and Y. Zhao, “A
multiobjective stochastic programming model for hydro-
power hedging operations under inexact information,” Water
Resources Management, vol. 31, no. 14, pp. 4649–4667, 2017.
[35] S. Séguin, S.-E. Fleten, P. Côté, A. Pichler, and C. Audet,
“Stochastic short-term hydropower planning with inflow
scenario trees,” European Journal of Operational Research,
vol. 259, no. 3, pp. 1156–1168, 2017.
[36] J. R. Birge and F. Loveaux, Introduction to Stochastic
Programming, Springer-Verlag, Berlin, Germany, 2nd edition,
2011.
[37] M. N. Hjelmeland, J. Zou, A. Helseth, and S. Ahmed,
“Nonconvex medium-term hydropower scheduling by sto-
chastic dual dynamic integer programming,” IEEE Trans-
actions on Sustainable Energy, vol. 10, no. 1, pp. 481–490, 2018.
[38] A. Helseth, M. Fodstad, and B. Mo, “Optimal medium-term
hydropower scheduling considering energy and reserve ca-
pacity markets,” IEEE Transactions on Sustainable Energy,
vol. 7, no. 3, pp. 934–942, 2016.
Advances in Advances in Journal of The Scientific Journal of
Operations Research
Hindawi
Decision Sciences
Hindawi
Applied Mathematics
Hindawi
World Journal
Hindawi Publishing Corporation
Probability and Statistics
Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 http://www.hindawi.com
www.hindawi.com Volume 2018
2013 www.hindawi.com Volume 2018

International
Journal of
Mathematics and
Mathematical
Sciences

Journal of

Hindawi
Optimization
Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

Submit your manuscripts at

www.hindawi.com

International Journal of
Engineering International Journal of
Mathematics
Hindawi
Analysis
Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

Journal of Advances in Mathematical Problems International Journal of Discrete Dynamics in

Complex Analysis
Hindawi
Numerical Analysis
Hindawi
in Engineering
Hindawi
Differential Equations
Hindawi
Nature and Society
Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

International Journal of Journal of Journal of Abstract and Advances in

Stochastic Analysis
Hindawi
Mathematics
Hindawi
Function Spaces
Hindawi
Applied Analysis
Hindawi
Mathematical Physics
Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018