Energy 307 (2024) 132601

Contents lists available at ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

Flexibility-based expansion planning of active distribution networks

considering optimal operation of multi-community integrated
energy systems
Chunling Wang a , Chunming Liu a, * , Xiulin Zhou a , Gaoyuan Zhang b
a
School of Electrical and Electronic Engineering, North China Electric Power University, Beijing, 102206, China
b
Key Laboratory of Power Station Energy Transfer Conversion and System (Ministry of Education), North China Electric Power University, Beijing, 102206, China

A R T I C L E I N F O A B S T R A C T

Handling editor: Neven Duic Large-scale integration of distributed renewable energy into active distribution networks leads to a dramatic
increase in flexibility requirements, and multi-community integrated energy systems may become important
Keywords: flexibility supplies owing to their multi-energy synergistic and complementary advantages. To fully utilize their
Active distribution network flexible regulation potential, a flexibility-based hierarchical expansion planning model for active distribution
Multi-community integrated energy system
networks considering optimal operation of multi-community integrated energy systems is proposed. In this
Expansion planning
model, the network and supply-demand flexibilities are integrated at the investment level to determine the
Renewable uncertainty
Insufficient flexibility risk installation types, locations, and capacities of distributed energy resources and power networks, and an inter-
action mechanism based on insufficient flexibility risk is designed at operating level to optimize operation
strategies and balance flexibility and economy. Considering the renewable uncertainties, a model-free scenario
generation method is introduced to capture the renewable seasonal and daily output characteristics without
relying on complicated explicit feature modeling. Finally, a distributed iterative solution method is devised based
on analytical target cascading theory. Simulations on a 54-node distribution system show the proposed approach
utilizes the flexibility of multi-community integrated energy systems to reduce the configuration needs of active
distribution networks for other flexible resources, improving the overall system economy and promoting
renewable consumption.

with REG volatility through electricity purchase and sale, thereby

reducing the configuration requirements of ADNs for other flexible re-
1. Introduction sources. However, MCIESs have their own energy supply methods, en-
ergy consumption characteristics, and stakeholders, their presence alters
With the vigorous promotion of energy conservation and emissions the composition structure of ADNs [4]. In this context, the expansion
reduction, active distribution networks (ADNs) that can support large- planning of restructured ADNs faces new requirements and challenges.
scale distributed renewable energy generators (REGs) have been Therefore, research on how to effectively utilize the flexible support
rapidly developed [1]. However, owing to the strong volatility and un- potential of MCIESs for ADNs and realize synergistic planning of REGs
certainty associated with REG outputs, ADNs must allocate sufficient and flexible resources is of considerable significance in postponing in-
controllable flexible resources to ensure efficient renewable consump- vestments and promoting efficient renewable resource consumption.
tion and reliable power supply [2]. Meanwhile, in the current context of
vigorously developing multi-energy coupling and flexible inter-
connecting technologies, an increasing number of community integrated 1.1. Literature review
energy systems (CIESs) are being integrated into ADNs owing to their
multi-energy complementarity and ladder utilization abilities [3]. For Several pioneering studies have investigated flexibility planning for
ADNs, multi-community integrated energy systems (MCIESs) have ADNs. Reference [5] presented a joint planning model for distributed
bidirectional regulation effects on electric power, and their numerous energy resources (DERs) and battery energy storage (BES) in ADNs to
internal controllable devices can function as flexible resources to cope facilitate the consumption of renewable energy. Reference [6] proposed

* Corresponding author.
E-mail address: [email protected] (C. Liu).

https://doi.org/10.1016/j.energy.2024.132601
Received 22 January 2024; Received in revised form 21 July 2024; Accepted 24 July 2024
Available online 26 July 2024
0360-5442/© 2024 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
C. Wang et al. Energy 307 (2024) 132601

Nomenclature cSi / cS0

i Construction and expansion costs of substations
cL / cTL Investment cost per unit length of newly built lines and
Abbreviations supporting tie-lines for SOP
ADN Active distribution network Grrated Unit installation capacity of type-r REG
REG Renewable energy generator zri,max Maximum installation quantity of type-r REG at candidate
CIES Community integrated energy system installation node i
MCIES Multi-community integrated energy system ∂ REG penetration rate
DER Distributed energy resources βBES Maximum magnification factor of BES
BES Battery energy storage Hgas Calorific value of natural gas
SOP Soft open point
ϕADN,u / ϕADN,d Permissible upper limits of CVaR for the upward and
GAN Generative adversarial network t t
downward insufficient flexibilities of ADN during period t
WGAN-GP Conditional Wasserstein generative adversarial network
with gradient penalty Variables
ATC Analytical target cascading f
Gri / Gi New installation capacities of type-r REG and type-f
Sets and Parameters flexible resource at node i
ΩR Set of candidate installation REG xSi / xS0
i Binary variable; it equals to 1 if installing newly built
ΩF Set of candidate installation flexible resources substation/expansion substation at node i
Λr / Λf Sets of candidate installation nodes of type-r REG/type-f xLij / xTL
ij Binary variable; it equals to 1 if installing newly built line/
flexible resource supporting tie-lines for SOP on line ij
ΛS / ΛS0 Sets of candidate installation nodes of newly built f
Pri,t / Pi,t Outputs of type-r REG and type-f flexible resource during
substations/expansion substations period t
ΛL Set of candidate installation nodes of newly built lines Pn,t
CIES,buy
Electricity purchased by ADN from the nth CIES
ΩM Set of typical scenarios
PCIES,sell Electricity sold by ADN to the nth CIES
ΩT Set of scheduling periods n,t
grid
ΩD Set of operating equipment in the AND Pt Electricity purchased by ADN from main grid
ΩN Set of CIES in the ADN Iij,t Current of line ij during period t
Ωd Set of operating equipment in CIES Uj,t Voltage magnitude at node j during period t
( )
b Discount rate Piin,t Hiin,t , Ciin,t Input powers of coupling devices in CIES
cr / cf Investment cost factors of type-r REG/type-f flexible
Hhes hes
dc,t / Hch,t Discharge and charge power of HES
resource

an ADN multistage distribution network planning method that consid- effective approach for describing uncertainty by transforming uncertain
ered active management of distributed generation units and demand factors into multiple deterministic scenarios. Scholars have commonly
response. In Ref. [7], a hierarchical multi-objective planning approach used statistical methods such as time series simulations [13], scenario
for ADNs is presented to improve the grid connection capabilities of tree generation [14], and the Copula function for generating REG output
photovoltaics (PVs) and wind turbines (WTs) by categorizing and scenarios [15]. Reference [16] utilized probabilistic statistical output
managing the demand-side response. An ADN optimization and plan- models for WT and PV to portray their uncertainties. A multi-stage
ning model with soft open point (SOP) was presented in Ref. [8] by scenario tree of intraday uncertainty was established in Ref. [17] to
considering the carbon emission cost. Reference [9] presented a dy- describe the stochastic output of REGs. A wind power scenario genera-
namic expansion planning model that was integrated with seasonally tion method based on Copula functions and forecast errors was pre-
transferred battery swapping stations and PVs. A coordination planning sented in Ref. [18]. However, REG outputs are affected by various
scheme of distributed generation, flexibility resource, and soft open complex factors such as geography, climate, and season. A primary
point integrated with energy storage system was proposed in Ref. [10] to challenge in statistical methods lies in the inability of simple models to
improve the flexible operation of active distribution networks. A flexible precisely describe the impact of multiple factors, which makes deriving
optimization planning model of distributed energy systems was pro- accurate model parameters difficult. With the rapid development of
posed in Ref. [11] that ensured system flexibility by fully accommoda- artificial intelligence, model-free scenario generation approaches are
ting the PV outputs under multiple scenarios. Reference [12] presented a gaining traction in power systems owing to their ability to learn data
distribution network planning method that considered the flexibility of distributions and generate new samples directly from historical data
virtual power plants, and utilized the upward and downward flexible without making assumptions about the distribution that the data obeys
adjustment of distributed resources to ensure the complete consumption [19]. The advantage over traditional statistical methods is the ability to
of PV outputs and load supply. Although these previous studies allocated mine the inherent distribution of uncertain variables, solve difficult
flexible resources such as BES, demand response, and SOP to improve modeling problems, and realize the unsupervised generation of sce-
the flexibility of distribution networks, fewer have focused on the flex- narios. In Ref. [20], a conditional style-based generative adversarial
ibility support ability of connected MCIESs, which have plentiful network (GAN) was applied to generate REG scenarios. In Ref. [21], a
controllable devices and incur no additional investment cost during GAN was leveraged to learn the spatio-temporal correlation of REG
ADN planning. In addition, most studies treated flexibility requirement outputs, and the Wasserstein distance was introduced to enhance model
as rigid constraints in planning models and lacked an evaluation of training quality. Reference [22] proposed a federated learning-based
insufficient flexibility risk, which may lead to overly conservative framework for renewable scenario generation which achieved privacy
planning schemes. protection.
The strong uncertainty of REG outputs significantly affects the stable Table 1 summarizes the main differences between the approach
operation and planning accuracy of distribution networks, making REG proposed in this study and the most relevant research in the field.
uncertainty modeling crucial for ADN planning. Scenario analysis is an Considering the above-mentioned previous studies, several research

2
C. Wang et al. Energy 307 (2024) 132601

Table 1
Comparison of the proposed approach with the related works.
Ref Planning subject REG and load uncertainties Modeling method

REG MT BES SOP Line Substation MCIES Handling method Insufficient flexibility
risk

[5] ✓ × ✓ × × × × Probability statistical × Bi-level iterative programming

method
[6] ✓ × × × ✓ ✓ × × × Single-level centralized optimization
[7] ✓ × × × ✓ × × × Insufficient power Bi-level iterative programming
supply
[8] ✓ × × ✓ × × × Probability statistical × Bi-level coordinated optimization
method
[9] × × ✓ × ✓ × × × × Single-level centralized optimization
[11] × × ✓ × × × × Scenario tree × Tri-level optimization
[12] × × ✓ × × × × Frank-Copula function REG curtailment rate Bayesian distributed robust optimization
This ✓ ✓ ✓ ✓ ✓ ✓ ✓ Data-driven method: CVaR of insufficient Analytical target cascading-based
paper CWGAN-GP flexibility distributed iterative optimization

gaps persist in flexibility planning for ADNs: (1) MCIESs have numerous The remainder of this paper is structured as follows: Section 2 de-
and diversified flexible controllable resources, yet their flexibility lineates the flexibility-based expansion planning framework of ADNs.
regulation potential in ADN planning has not been fully explored; (2) The model-free scenario generation model based on CWGAN-GP is
The risk of insufficient flexibility is exacerbated with a high percentage introduced in Section 3. Sections 4 and 5 detail the bi-level expansion
of REGs, but previous studies have typically neglected to analyze or planning model and its solution approach, respectively. Section 6
quantify this risk in-depth, instead adopted rigid constraints of flexi- demonstrates the effectiveness of the proposed method through a case
bility requirements leading to overly conservative planning schemes or study. Finally, Section 7 provides a summary of the study.
redundant flexible resources; (3) Most scenario generations are
commonly modeled using statistical methods that require fitting specific 2. Flexibility-based expansion planning framework
model parameters. However, modeling becomes extremely difficult
when diversity, correlation, and seasonality of renewable energy are The flexibility of ADNs is reflected in the supply-demand balance of
considered simultaneously. node flexible resources and the smoothness of network channels. The
node flexible resources mainly include DERs such as microturbines
(MTs), BESs, and substations. The network-side resources include the
1.2. Contribution transmission lines and inter-nodal interconnections represented by
SOPs. The proposed ADN hierarchical expansion planning framework
To address the aforementioned gaps, this study presents a flexibility- considering the optimal operation of MCIESs is illustrated in Fig. 1, and
based expansion planning framework for ADNs that effectively utilizes consists of the following parts:
the operational flexibility of MCIESs in uncertain environments.
Compared with the exiting research, the main contributions of this paper (1) Considering the long-term seasonal variations in load demands,
are as follows: three typical seasonal (winter, summer, and transition) daily
loads of ADN and MCIESs are selected. Next, considering the
(1) To fully utilize the flexibility of MCIESs, a flexibility-based ADN long- and short-term uncertainties of REGs, the CWGAN-GP sce-
expansion planning method considering the optimal operation of nario generation model is constructed to generate numerous REG
MCIESs is proposed. This method can utilize the complementary output scenarios with seasonal and daily fluctuation character-
coordination among multiple energy sources in MCIES operation istics of historical data, and scenario reduction is utilized to
to reduce the total planning cost, as well as improve the supply- obtain typical scenario sets.
demand and network flexibilities; Moreover, an ADN-MCIES (2) The ADN-MCIES coordinated operation level, which considers
interaction mechanism considering the risk of insufficient flexi- ADN-MCIES interaction mechanism based on insufficient flexi-
bility is developed to reduce the uncertainty risk of REG outputs bility risk is constructed to optimize their operation strategies by
and strike a balance between flexibility and economy. adopting the scenario-based stochastic optimization. The ADN
(2) To address the long- and short-term uncertainties of REG outputs, and MCIESs iteratively operate with tie-line power as the transfer
a model-free scenario generation approach based on the condi- parameter until their interaction schemes are consistent. The
tional Wasserstein generative adversarial network with gradient structure and energy flow of a typical CIES investigated in this
penalty (CWGAN-GP) is employed, which effectively captures the study are shown in Fig. 2.
seasonal and daily output characteristics of REGs and avoid (3) The expansion planning model that integrates network and
complex explicit feature modeling; supply-demand flexibilities for ADNs is built, in which the
(3) A distributed iterative solution method based on the analytical installation types, locations, and capacities of DERs and power
target cascading (ATC) theory is designed to simulate the game networks are optimized at the investment level by minimizing the
process among all parties and efficiently solve in parallel using annual investment cost and operating cost transmitted from the
only necessary data to be transmitted, which effectively protects operation level. The operational management of DERs and power
the privacies of all parties and ensures convergence of the networks is carried out according to the established ADN-MCIES
solution. optimal scheduling model by adopting the installation strategy
(4) A case study is undertaken to validate the proposed methods and derived from the investment level.
showcase the revenue potential of MCIES engaging in ADN
planning using multi-energy coupling. The proposed method
mitigates the penalty cost associated with insufficient flexibility
and reduces the total cost encompasses investment and operation
phases.

3
C. Wang et al. Energy 307 (2024) 132601

Fig. 1. ADN hierarchical expansion planning framework.

authenticity of the generated samples is achieved through mutual game

learning between these two networks. Conditional generative adversa-
rial networks (CGANs) extend the traditional GAN by incorporating
conditional information while retaining the game structure of the
generator and discriminator. This augmentation empowers the gener-
ator to comprehend the mapping relationship of the sample probability
distribution for specific conditions, leading to a substantial generaliza-
tion effect on the specified type of data samples [24]. The network
structure of the CGAN is depicted in Fig. 3.
As illustrated in Fig. 3, the noise, z, and condition, y, are combined as
the inputs of the generator, and the outputs are the generated samples,
xʹ. The discriminator processes real samples, x, conditions, and gener-
ated samples as inputs, yielding discrimination results. In this study,
sample labels are utilized to generate renewable scenarios, considering
Fig. 2. Overall structure and energy flow of a typical CIES. various seasonal characteristics (i.e., winter, summer, and transition) as
conditions.
3. Conditional Wasserstein generative adversarial network- The goal of the discriminator in the CGAN is to reduce the discrim-
based scenario generation inative value of the generated scenarios, enhance that of the real sce-
narios, and ascertain whether the generated scenarios fulfill the
The GAN, a deep learning model introduced by Goodfellow, com- specified conditions. Simultaneously, the generator strives to generate
prises two key components: the generator G and the discriminator D scenario samples that adhere to these conditions and elevate the
[23]. The generator utilizes a black box pattern instead of a statistical discriminative value of the generated scenarios. Both networks are
model to avoid complex modeling of explicit features. Enhanced trained to achieve Nash equilibrium through a zero-sum game, enabling

Fig. 3. Basic structure of CGAN.

4
C. Wang et al. Energy 307 (2024) 132601

the generator to effectively map corresponding conditional scenarios. 4. Hierarchical expansion planning model
Therefore, the objective function of the CGAN is [20]:
Based on the flexibility expansion planning framework proposed in
minmaxV(D, G) = Ex∼P(x) [log D(x|y)] + Ez∼P(z) [log(1 − D(xʹ))] (1)
G D Section 2, the ADN flexibility expansion planning model considering the
optimal operation of MCIESs is constructed.
where, P and E represent the probability distribution and its expected
value; D( ) and G( ) are the outputs of the discriminator and generator.
To address the issues of instability and pattern collapse during CGAN 4.1. Investment level
training, the Wasserstein distance is utilized in the discriminator loss
function to measure the probability distance between the real and The minimum total cost composed of the annualized investment cost
generated scenarios [25]. CADN ADN
inv and annualized operating cost Cope is adopted as the objective
function to be optimized as follows:

min CADN = CADN inv + Cope

ADN

∑∑ ∑∑ ∑ ∑ ∑ ∑
CADN
inv = τr cr Gri + τf cf Gfi + τS cSi xSi + τS0 cS0
i xi +
S0
τL cL xLij Lij + τTL cTL xTL
ij Lij
r∈ΩR i∈Λr f ∈ΩF i∈Λf i∈ΛS i∈ΛS0 ij∈ΛL ij∈ΛSOP (4)
∑
CADN
ope = 365 ppro,m CADN
m,ope
m∈ΩM

( )
W Px, Pxʹ = sup EP(x) D(x|y) − EP(xʹ) D(xʹ|y) (2) where, τ = b(1 + b)σ /[(1 + b)σ − 1] is the coefficient for converting the
‖fD ‖L ≤1 present value of equipment to its annual value, in which b and σ are the
discount rate and lifespan, respectively; ΩR is the set of candidate
where, ‖fD ‖L ≤ 1 indicates that the discriminator function needs to installation REGs, including PV and WT; ΩF is the set of flexible re-
satisfy the 1-Lipschitz continuum. sources including MT, BES, and SOP; Λr , Λf , ΛS , ΛS0 , ΛL , and ΛSOP are
In addition, a gradient penalty is introduced to make the discrimi- sets of candidate installation nodes for REGs, flexible resources, newly
nator approximate the 1-Lipschitz continuum, which results in the built substations, expansion substations, newly built lines, and sup-
following objective function of CWGAN-GP [26]:
porting tie-lines for SOP, respectively; cr and cf are the investment cost
minmaxV(D, G) = Ex∼P(x) [log D(x|y)] + Ez∼P(z) [log(1 − D(xʹ))] factors of REG and flexible resources; cSi and cS0
i are the construction and
G D
[ ] expansion costs of substations; cL and cTL are the investment cost per unit
x |y)‖ − 1)2
− λEx̂∼Px̂ (‖∇x̂ D( ̂ (3) f
length of newly built lines and supporting tie-lines for SOP; Gri and Gi are
the new installation capacities of type-r REG and type-f flexible re-
where, λ is the penalty factor; ̂
x = εx + (1 − ε)G(z), and ε obeys a uni- sources at node i, respectively; xSi and xS0
i are 0–1 decision variables for
form distribution over [0,1]. whether newly built substations and expansion substations are installed
The training process of CWGAN-GP mainly consists of the following
at node i; xLij and xTL
ij are 0–1 decision variables for whether newly built
steps:
lines and supporting tie-lines for SOP are installed on line ij; Lij is the
Step 1. Normalize the training samples and construct high- length of line ij; ΩM denotes the set of typical scenarios; ppro,m and CADN
m,ope
dimensional noises conforming to the standard normal distribution, are the probability and daily operating cost of scenario m, respectively.
splice the noises with the condition values corresponding to the training The investment constraints include:
samples vertically and input them into the generator, which outputs the
generated samples. (1) REG installation capacity and quantity constraints
Step 2. Splice the condition values with the training and generated Gri = zri Grrated
samples vertically, respectively, and then input them into the discrimi-
nator, which outputs the discriminative values for the real and gener- 0 ≤ zri ≤ zri,max
∑∑ (5)
ated samples. Gri ≤ ∂PLmax
r∈ΩR i∈Λr
Step 3. Calculate the loss functions of the generator and discriminator,
and use the RMSprop optimization algorithm to update the weights of where, Grrated is the unit installation capacity of type-r REG; zri,max is the
the generator and discriminator networks respectively. If the training is
maximum installation quantity of type-r REG at candidate installation
not finished, return to Step 1 for the next round of training.
node i; ∂ is the REG penetration rate; PLmax is the maximum load demand.
Step 4. After training, extract the generator model from CWGAN-GP,
input fixed condition values and n high-dimensional noises that (2) Flexible resource installation capacity and quantity constraints
conform to normal distribution, and the generator would output n
renewable generation output scenarios that meet the condition values. The installation capacity and rated power constraints of BES are as
follows:
Upon generating numerous renewable scenarios for three seasons
using CWGAN-GP, typical scenarios are obtained using K-Medoids and 0 ≤ GBES
i ≤ GBES
i,max
then applied to the operation level of the hierarchical planning model 0 ≤ SBES ≤ SBES (6)
i i,max
[27].
GBES
i ≤ SBES
i βBES

5
C. Wang et al. Energy 307 (2024) 132601

where, GBES BES

i,max and Si,max are the maximum installation capacity and rated
sources for ADN when its upward or downward flexibility is insufficient.
power of BES at node i, respectively; βBES is the maximum magnification
factor. 4.2.1. Flexibility interaction mechanism
The installation capacity constraint of MT is given as: Owing to the volatility and uncertainty of REGs and load demand,
insufficient flexibility risk exists in each CIES and ADN. Scientific
0 ≤ GMT
i ≤ GMT
i,max (7) assessment of this risk is useful for assisting decision-makers in choosing
a rational planning scheme. Therefore, based on the conditional value at
where, GMT
i,max is the maximum installation capacity of MT at node i. risk (CVaR) theory in economics [30], a CVaR of insufficient flexibility is
The installation capacity and quantity constraints of SOP are [28]: proposed to quantify the risk and discretized by combining typical
scenarios as follows:
GSOP
ij = zSOP
ij Grated
SOP
(8) 1 ∑
0 ≤ zSOP
ij ≤ zSOP
ij,max xij
TL
Fren,t
CVaR
= α1 + ppro,m PCR
1 − β m∈ΩM t

(11)
where, GSOP SOP
rated is the unit installation capacity of SOP; zij,max is the 1 ∑
Fload,t
CVaR
= α2 + ppro,m PCL
maximum installation quantity of SOP on line ij. 1 − β m∈ΩM t

(3) Network topology constraints CVaR

where, Fren,t CVaR
and Fload,t are the CVaRs of renewable curtailment and load
shedding, respectively, during period t (i.e., the expected value that
Given that the candidate installation locations for SOPs are at the exceeds the flexibility deficit threshold at a given confidence level β); α1
interconnection switches, coordinated planning necessitates the estab- and α2 are the boundary values of renewable curtailment and load
lishment of a radiation network and tie-lines. To ensure that the tie-lines
shedding; PCR CL
t and Pt are the powers of renewable curtailment and load
constructed for SOP installation do not intersect with the selected lines
shedding, which reflects the system supply-demand flexibility.
of the radial network, a pertinent logical constraint is introduced, as
Based on the CVaR of insufficient flexibility, the following ADN-
follows:
MCIES flexibility interaction mechanism is designed:
xTL
ij + xij ≤ 1
L
(9)
(1) Flexible resources are supplied by CIES to ADN when the CIES
A restructured ADN needs to maintain a radial and connected grid flexibility is sufficient and the ADN flexibility is insufficient;
structure without loops or islands. The following equations are then used (2) Flexible resources are provided by ADN to CIES when the ADN
for constraints [29]: flexibility is sufficient and the CIES flexibility is insufficient;
Wij + Wji = δij i, j = 1, ..., q (3) The CIES and ADN perform flexibility interaction based on eco-
∑ nomics when both flexibilities are sufficient;
Wij = 1 i = 2, ..., q (10)
j∈ΩL (i)
(4) The CIES and ADN perform flexibility interaction based on the
W1j = 0 j ∈ ΩL (1) CVaR of insufficient flexibility when both flexibilities are
insufficient.
where, δij is the 0–1 variable that describes the on-off state of branch ij,
with 0 being disconnected and 1 being closed; q is the number of ADN 4.2.2. Optimal scheduling of active distribution network
nodes; Wij is a 0–1 variable, where 1 indicates that the current of branch The ADN operating level manages the operation of the DERs and
ij flows from node j to node i, and 0 indicates the opposite; ΩL (i) is the set power network with the optimization goal of minimizing the operating
of nodes connected to node i. cost for each typical day. The scenario m subscript is omitted owing to
the consistent scheduling model for each typical scenario.

4.2. Operating level

min CADN
ope = C1 + C2 + C3 + C4 + C5
∑[ ( MT / MT / )]
C1 = γ ADN
gas Pt η Hgas Δt
t∈ΩT
⎛ ⎞
∑ ∑ ∑ ∑ ∑ ∑∑ ∑∑ f
C2 = κi xi +
S
κi xi +
S0
κij xij +
L
κij xij +
TL ⎝ κi Pi,t +
r
κi Pi,t ⎠Δt
i∈ΛS i∈ΛS0 ij∈ΛL ij∈ΛSOP t∈ΩT r∈ΩR i∈Λr f∈ΩF i∈Λf
∑∑( ) (12)
C3 = t Pn,t
γbuy CIES,buy
t Pn,t
− γ sell CIES,sell
Δt
t∈ΩT n∈ΩN
∑
C4 = t Pt Δt
γgrid grid

t∈ΩT
[ ]
∑ ∑ ∑( )
C5 = Pr,E
i,t − Pri,t CR CL CL
γ + Pt γ Δt
t∈ΩT r∈ΩR i∈Λr

An ADN-MCIES flexibility interaction mechanism and coordinated where, C1 , C2 , C3 , C4 , and C5 are the fuel cost of MT, operation and
optimization scheduling models of ADN and MCIESs are constructed, so maintenance (O&M) cost of substations, lines, REGs, and flexible re-
that MCIESs can provide upward or downward flexibility adjustment sources, interaction cost with MCIESs, power purchase cost from the
through the coupling and complementarity among multiple energy main grid, and insufficient flexibility penalty cost, respectively; γADN
gas is

6
C. Wang et al. Energy 307 (2024) 132601

the price of natural gas, and Hgas is the calorific value of natural gas; PMT
t
(4) SOP power constraints
is the MT output during period t; ηMT is the electrical efficiency of MT; κi
is the O&M cost factor of equipment i; Pri,t and Pi,t are the outputs of REGs
f SOPs are generally back-to-back voltage source converters consisting
of high-power, fully-controlled power electronic components [31]. The
and flexible resources during period t; ΩN is the set of CIESs in the ADN;
buy CIES,buy
steady state control mode of PQ-VdcQ is employed in this study, i.e., one
γt and Pn,t are the price and power of electricity purchased by the converter realizes stable control of the DC voltage and the other con-
ADN from the nth CIES, respectively; γ sell
t and PCIES,sell
n,t are the price and verter realizes flexible control of the transmitted power [32]. Although
electricity sold by the ADN to the nth CIES, respectively; γt
grid
and Pt
grid
are SOP has certain losses, it can be ignored relative to the entire distribu-
tion system. Therefore, the power output constraints of SOPs are:
the price and electricity purchased by the ADN from the main grid; Pr,E
i,t
and Pri,t are the expected output and consumption of REGs during period PSOP
i,t + Pj,t = 0
SOP

CR CL √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
)2̅
t; γ and γ are the penalty cost factors of renewable curtailment and ( )2 (
load shedding; Δt is the scheduling interval; and ΩT is the set of PSOP
i,t + QSOP i,t ≤ PSOP
ij,max (16)
scheduling periods. √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
( ) 2 ( 2) ̅
The constraints for optimal scheduling of the ADN are described PSOP
j,t + QSOP
j,t ≤ PSOP
ij,max
below.
where, PSOP SOP SOP SOP
i,t , Qi,t and Pj,t , Qj,t are the active and reactive outputs of
(1) AC power flow constraints
SOP at nodes i and j, respectively; PSOP
ij,max is the maximum active power of
∑ ∑ ( ) ∑ ( )
∑(
Pjk,t − Pij,t − Iij,t
2
Rij = Pdj,t − PLj,t − PCL SOP on line ij.
j,t + PCIES,buy
n,t
k∈ΩL (j,:) i∈ΩL (:,j) d∈ΩD n∈ΩN
) ∑ ∑ ( ) (5) Flexibility interaction power constraints with MCIES
− PCIES,sell
n,t Qjk,t − Qij,t − Iij,t
2
Xij
k∈ΩL (j,:) i∈ΩL (:,j) 0 ≤ PCIES,buy
n,t ≤ δCIES,buy
n,t PCIES,buy
n,max
( ) ( ) P2ij.t + Q2ij,t 0 ≤ PCIES,sell ≤ δCIES,buy PCIES,sell (17)
= QLj,t U2j,t = Ui,t
2
− 2 Rij Pij,t + Xij Qij,t + R2ij + Xij2 Iij,t
2 2
Iij,t = n,t n,t n,max
U2i,t δCIES,buy + δCIES,sell ≤1
n,t n,t
(13)
CIES,buy
where, δn,t and δCIES,sell
n,t are 0–1 variables describing the purchasing
where, ΩL (j, :) is the set of end nodes of branches with j as the first node,
CIES,buy
and ΩL (:, j) is the set of first nodes of branches with j as the end node; ΩD and selling status of the ADN from the nth CIES, respectively; Pn,max
is the set of operating equipment in the ADN, including REGs, MTs, BESs, and PCIES,sell
n,max are the maximum power purchased and sold by the ADN
SOPs, and substations; Pjk,t , Pij,t and Qjk,t , Qij,t are the active and reactive from the nth CIES during period t, respectively.
powers of lines jk and ij during period t; Rij and Xij are the resistance and
reactance of line ij; Pdj,t is the active power injected into node j during (6) Insufficient flexibility risk constraints
period t by operating equipment d; PLj,t and QLj,t are the original active and Fren,t
ADN,CVaR
≤ ϕADN,u
t
reactive loads at node j; Iij,t is the current of line ij; and Uj,t is the voltage ADN,CVaR
(18)
Fload,t ≤ ϕADN,d
t
magnitude at node j during period t.

where, ϕADN,u
t and ϕADN,d
t are the permissible upper limits of the CVaR for
(2) Security constraints
the upward and downward insufficient flexibilities of ADN during
Ui,min
2
≤ U2i,t ≤ Ui,max
2
period t.
(14) In addition, the MT and BES operating constraints as well as the REG
0 ≤ Iij,t
2
≤ Iij,max
2
consumption and load shedding constraints are commonly used and thus
where, Ui,min and Ui,max are the minimum and maximum values of the not discussed in this paper; the details can be found in our previous work
node voltage, respectively; Iij,max is the maximum current of line ij. [21].

4.2.3. Optimal scheduling of multi-community integrated energy system

(3) Substation power constraints
The MCIES operating level optimizes the operation strategies of
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
( ) 2 ( ) 2 REGs and controllable devices in each CIES by minimizing the operating
PSi,t + QSi,t ≤ xSi GSi,rated , ∀i ∈ ΩS
cost for each typical day. Similar to the ADN scheduling model, the
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (15)
( ) 2 ( ) 2 scenario m subscript is omitted.
S0 S0,new
PS0
i,t + QS0
i,t i,rated + xi Gi,rated , ∀i ∈ ΩS0
≤ GS0

[ ]
∑ ( / / ) ∑ ( tse ) ( cr cr ) ( sell ADN,buy )
min CCIES
ope = γgas Pmt
t
mt
η Hgas + Pt κ + γ × min Pt , 0 + γ Pt + γ Pt + γ t Pt
i i dr cl cl
− γt Pt
buy ADN,sell
Δt (19)
t∈ΩT i∈Ωd

where, γgas is the price of natural gas purchased by CIES; Pmt

t and η
mt
are
where, PSi,t , QSi,t and PS0 S0
i,t , Qi,t are the active and reactive power outputs of
the output and electrical efficiency of MT in CIES; Ωd is the set of
newly built and expansion substations; GSi,rated is the rated capacity of
operating equipment in CIES, including REG, MT, heat recovery boiler
newly built substations; GS0
i,rated is the rated capacity of existing sub- (HRB), electric chiller (EC), absorption chiller (AC), electric boiler (EB),
stations; GS0,new
i,rated is the rated expansion capacity of existing substations. BES, heat energy storage (HES), and cooling energy storage (CES); Pit is

7
C. Wang et al. Energy 307 (2024) 132601

the output of operating equipment i during period t; γ dr is the demand conversion and iterative algorithm are presented to facilitate the solving
response compensation cost factor, and Ptse
t is the time-shiftable electric
process.
load during period t; Pcr cl
t and Pt are the renewable curtailment and load
shedding powers in CIES during period t; Pt
ADN,buy
and PADN,sell are the 5.1. Hierarchical model conversion
t
electricity purchased and sold by CIES from the ADN, respectively.
The constraints for CIES optimal scheduling include: Considering that ADN and MCIES belong to different stakeholders,
the interactive iteration between ADN and MCIES is retained, and the
(1) Power balance constraints investment and operating levels of ADN are converted into a single-level
⎧ model through variable replacement to avoid complicated parameter
⎪ Pt +Pt +Pt +Pdc,t +P,t
⎪
⎪ wt pv mt bes ADN.buy
= Plt +Ptse
t +Pch,t +Pin,t +Pin,t − Pt +Pt
bes eb ec cl ADN,sell passing.
⎨
Hout,t +Hout,t +Hdc,t = Ht +Hch,t +Hin,t
rb eb hes l hes ac In terms of REGs, the parameter Pr,E
i,t in Eq. (12) is replaced using the
⎪
⎪ following equation:
⎪
⎩ Cac +Cec +Cces = Cl +Cces
out,t out,t dc,t t ch,t

(20) PWT,E
i,t = GWT WT,E
i Punit,i,t
(23)
PPV,E
i,t = GPV PV,E
i Punit,i,t
where, Pwtt and Ppv
t are the consumption power of WT and PV during
period t; Plt , Hlt , and Clt are the electric, heat, and cooling loads, respec- where, PWT,E PV,E
unit,i,t and Punit,i,t are the per unit expected outputs of WT and PV.
( ) ( )
tively; Piin,t i
Hin,t , Ciin,t and Piout,t Hiout,t , Ciout,t are the input and output For flexible resources, the rated power of MT, BES, and SOP in the
ADN operational constraints are replaced with:
powers of coupling devices in CIES; Pbes bes hes hes ces
dc,t (Pch,t ), Hdc,t (Hch,t ), and Cdc,t
(Cces
ch,t ) are the discharge and charge powers of BES, HES, and CES,
PMT
i,max = Gi
MT

respectively. PBES
i,max = Gi
BES
(24)
PBES
ij,max = Gij
SOP

(2) Flexibility interaction power constraints with ADN

0 ≤ PADN,buy
t ≤ δADN,buy
t PADN,buy
max 5.2. Analytical target cascading-based distributed solving
0 ≤ PtADN,sell
Pmax
ADN,sell ADN,sell
≤ δt (21)
δADN,buy
t + δADN,sell
t ≤1 The iterations between ADN and MCIES are addressed by leveraging
the ATC algorithm. ATC, a recent mathematical optimization method
where, δt
ADN,buy
and δADN,sell are 0–1 variables describing the purchasing known for its superior convergence, couples systems through decision
t
ADN,buy variables in hierarchical structures [36]. A typical ATC structure is
and selling status of the CIES from the ADN; Pmax and PADN,sell
max are the
shown in Fig. 4, in which O1 is the main system, S1,1 -S1,n are
maximum power purchased and sold by CIES from the ADN during
sub-generation systems, S2,1 -S2,j and S2,k -S2,m are second-generation
period t, respectively.
systems. Each system can set its own optimization model, where the
variables contained in the optimization model of each level that are
(3) Insufficient flexibility risk constraints
common with other levels are coupled variables. The information flow is
Fren,t
CIES,CVaR
≤ ϕCIES,u
t
transmitted level by level from top to bottom, where the lower level
CIES,CVaR
(22) system optimizes its own performance by moving closer to the coupled
Fload,t ≤ ϕCIES,d
t
variables passed from the upper level system while satisfying
self-constraints. The optimized coupled variables of the lower level
where, ϕCIES,u
t and ϕCIES,d
t are the permissible upper limits of the CVaR for
system are then fed back as responses to the upper level system, which
the upward and downward insufficient flexibilities of CIES during
receives them as constants and adds them to its objective for optimiza-
period t, respectively.
tion, similar to the lower level systems. This loop continues until the
Other operating constraints in each CIES, including the power output
conditions are satisfied.
and ramp rate constraints of MT and multi-energy coupling equipment
In the proposed model, the ADN and MCIES are coupled through tie-
[33], capacity constraints of BES, HES, and CES [34], as well as the REG
line powers. Following the initial optimal dispatch of the ADN, the
consumption and load shedding constraints [35], are commonly used, ADN
and are thus not discussed in this paper. optimized tie-line power, Pn,t , is passed on as a known parameter to the
MCIES. Subsequently, a penalty function is added to the objective
5. Model solving function of each CIES to reduce the deviation of the interaction power
ADN
from Pn,t as follows:
The ADN hierarchical expansion planning model constructed above
T ⃒ ⃒
contains interactive iterations between the investment and operating ∑ ⃒ ADN ⃒
min CCIES
n,ope + ωn,t ⃒PADN,buy
n,t − PADN,sell
n,t − Pn,t ⃒ (26)
levels, as well as between the ADN and MCIES, which considerably in- t=1
crease the solving complexity. Therefore, the following model
where, ωn,t is the penalty function multiplier of CIES n during period t.
Similarly, after the optimization of each CIES is completed, the tie-
CIES
line power, Pn,t , is transmitted to the ADN. A penalty function is also
added to the objective function of the ADN to reduce the deviation of its
CIES
interaction power from Pn,t :
T ∑
∑ N ⃒ ⃒
⃒ CIES ⃒
min CADN + ωn,t ⃒PCIES,buy
n,t − PCIES,sell
n,t − Pn,t ⃒ (27)
t=1 n=1

This process is repeated until the following convergence criteria are

Fig. 4. Typical ATC system structure. satisfied:

8
C. Wang et al. Energy 307 (2024) 132601

Table 2 ⃒ ( )⃒
⃒ ⃒
ADN parameters. ωn,t,k = ωn,t,k− 1 + χ n,t,k− 1 ⃒PADN,buy
n,t,k − PADN,sell
n,t,k − PCIES,buy
n,t,k − PCIES,sell
n,t,k ⃒
(29)
Parameter Value Parameter Value Parameter Value χ n,t,k = λχ n,t,k− 1

b 0.08 Hgas (kWh/m3) 9.7 ηMT 0.44

σ (year) 20 zrmax 50 κWT ($/kW) 0.0064 where, χ n,t,k is the first-order term multiplier of the kth iteration during
cWT ($/kW) 814 ∂ 0.7 κPV ($/kW) 0.0044 period t; λ is the iterative multiplier, which is generally in the range of
cPV ($/kW) 525 GBES
max (kWh)
2000 κBES ($/kW) 0.01 2 < λ < 3. Since only the tie-line powers and transaction price need to
cMT ($/kW) 202 SBES
max (kW)
500 κMT ($/kW) 0.0058 be transmitted at each iteration, the privacy of each stakeholder is
cBES ($/kW) 218 βBES 0.25 κL ($) 436 effectively protected.
cSOP ($/kW) 145 GMT
max (kW)
10000 γCR ($/kW) 0.1
cL ($/km) 35650 Grrated (kW) 100 γCL ($/kW) 0.2
cTL ($/km) 35650 GSOP 100 ϕADN,u (kW) 500 6. Case study
rated (kW)
γGAS ($/m3) 0.4 zSOP
max
50 β 0.95
The proposed expansion planning approach is tested on a 54-node
distribution system, where the candidate installation nodes for WTs,
Table 3
PVs, MTs, and BESs are 12, 14, 17, 24, 39, and 47; 5, 8, 10, 26, 30, 33,
MCIES parameters. and 37; 9, 13, 16, and 32; 8, 9, 13, 16, 17, 32, and 50, respectively. The
candidate newly built lines for SOPs are 22–23, 6–28, 30–43, 33–39,
Parameter Residential CIES Commercial CIES Industrial CIES
46–47, 2–8, and 10–15. The residential CIES, commercial CIES, and
Pmax
mt (kW) 1000 650 1400 industrial CIES are connected to nodes 4, 40, and 20, and it is assumed
Pmin
mt (kW)
200 100 200
that the loads in the above CIESs are represented by residential, com-
rdmt / rumt (kW) 500 300 700
mercial, and industrial loads, respectively. The CWGAN-GP model is
Pmax
eb (kW) 400 400 900
200 200 400
built via the deep learning framework Pytorch, the measured REG
rdeb / rueb (kW)
Pmax output data from the Elia integration datasets from June 1, 2021 to May
ac (kW) 500 500 800
Pmin
ac (kW)
60 60 80 31, 2023 are selected as the sample set, and the continuous data of one
rdac / ruac (kW) 200 200 400 day are taken as one-day samples, from which 80 % of the samples are
Pmax
ec (kW) 100 200 500 randomly selected for training and the remaining 20 % are taken as the
rdec / ruec (kW) 50 100 200 testing samples. The learning rate and batch size in training are 0.0002
ϕuCIES / ϕdCIES 200 200 500 and 64, respectively. The substation parameters are shown in Ref. [37],
(kW)
the ADN parameters are listed in Table 2, the MCIES parameters are
γgas ($/M3) 0.45 0.55 0.55
listed in Table 3 and [33], and the time-of-use (TOU) and main grid
βCIES 0.95 0.95 0.95
prices are shown in Fig. 5. The initial values of χ n and ωn are 0.0001 and
0.02, ε1 and ε2 are set to 0.01, T = 24 h, and Δt = 1 h. The planning
procedure is performed using MATLAB R2018a with the CPLEX 12
solver.
To verify the effectiveness and superiority of the proposed model, the
ADN planning with and without consideration of optimal MCIES oper-
ation is compared. In addition, SOPs, as flexible power electronic de-
vices connecting different power supply areas of an ADN, demonstrated
a significant impact on the network flexibility. Consequently, four ADN
planning modes are compared in this study. Mode 1 represents the
conventional planning method that does not consider MCIES optimal
operation and SOP access. Mode 2 enhances M1 by incorporating MCIES
optimal operation into the ADN planning. Mode 3 improves upon Mode
1 by configuring SOPs. Mode 4 i.e. the proposed model combines MCIES
optimal operation and SOP access in ADN planning.

Fig. 5. TOU price and main grid price.

6.1. Scenario generation analysis
⃒ ( )⃒
⃒ ADN,buy ⃒
⃒Pn,t,k − PADN,sell − PCIES,buy − PCIES,sell ⃒ ≤ ε1
n,t,k n,t,k n,t,k The effectiveness of scenario generation is analyzed in terms of
⃒(
⃒ ADN ∑ N
) (
∑N
)⃒
⃒ seasonal characteristics, diversity and correlation, as well as typical
⃒ Ck + CCIES
n,k − C ADN
k− 1 + C CIES
n,k− 1
⃒ (28) daily scenarios of REGs.
⃒ n=1 n=1
⃒
⃒ ( ) ⃒ ≤ ε2
⃒ N
∑ ⃒
⃒
⃒ C ADN
k− 1 + CCIES
n,k− 1
⃒
⃒ 6.1.1. Seasonal characteristics statistical analysis
n=1
To demonstrate the effectiveness of the CWGAN-GP scenario gen-
where, ε1 and ε2 are convergence accuracies. Equation (28) indicates eration method in portraying the seasonal output characteristics of
that in the kth iteration, the tie-line powers of the ADN and MCIE are as REGs, a comparison is made between the average daily generation of
coupled variables and their difference should satisfy the accuracy real and generated samples of the WT and PV plants across different
requirement. In addition, the fluctuation of the total costs of the ADN seasons. This comparison is depicted in Fig. 6.
containing MCIESs in the kth and k-1st iterations should satisfy the ac- As shown in Fig. 6, the average daily generation of the generated
curacy requirement, thus ensuring the overall benefit of the ADN. samples under different seasons is similar to that of the historical sam-
If the convergence criteria cannot be satisfied simultaneously, the ples and captured the significant seasonal difference—WTs generate
penalty function multipliers are updated according to Eq. (29). more electricity in winter and less in summer, while PV generate more
electricity in summer and less in winter. These results demonstrate that
the proposed CWGAN-GP-based scenario generation can accurately
capture the seasonal output characteristics of REGs and effectively track

9
C. Wang et al. Energy 307 (2024) 132601

Fig. 6. Average daily power generation of (a) WT and (b) PV under different seasons.

Fig. 7. Pairs of generated and real samples of REGs and their autocorrelation coefficients.

Fig. 8. Seasonal typical scenarios of REGs.

their seasonal output trends. Fig. 7(a) shows that the generated samples effectively capture the
diversity of modal changes. However, the pairs of samples do not
6.1.2. Diversity and correlation analysis completely coincide, which demonstrates that the scenario generation
An output comparison of sample pairs of REGs selected from the model performs fitting instead of simply memorizing the probability
generated and real samples via the Euclidean distance nearest principle distribution of the historical data. This indicates that the CWGAN-GP
and their autocorrelation coefficients are shown in Fig. 7. model is trained without overfitting, and the generated data capture

10
C. Wang et al. Energy 307 (2024) 132601

Table 4 lower in winter, higher in summer, and at intermediate levels in the

Typical daily scenario probabilities. transition season, whereas the WT outputs are the opposite and more
Typical Scenario probability volatile and diverse, in line with the seasonal output characteristics of
scenario REGs. This indicates that the clustering scenarios effectively reflect the
Typical Typical Typical Typical
scenario 1 scenario 2 scenario 3 scenario 4 seasonal and daily fluctuation characteristics of the REGs.

Winter 0.0789 0.0493 0.0838 0.0345

Summer 0.0832 0.0555 0.0504 0.0630
6.2. Expansion planning scheme analysis
Transition 0.1153 0.0953 0.1504 0.1404
The expansion planning results and schemes under the four modes
are presented in Fig. 9 and Table 5, respectively.
the characteristics of the historical data, such as rapid changes, spikes, As shown in Fig. 9 and Table 5, the proposed model realizes the
and daily cycle changes. Fig. 7(b) shows that the autocorrelation co- coordinated expansion planning of REGs and diversified flexible re-
efficients between the generated and real samples gradually decrease as sources in ADNs. Among them, two new substations are built in Modes 1
the time interval increases and all trends are basically the same, indi- and 2 that do not consider SOPs. In contrast, as SOPs are considered in
cating that the generated samples have temporal correlation features Modes 3 and 4, only one new substation is built in the ADN despite the
similar to those of the historical samples. Therefore, the generated sce- increase in newly built lines. This is because SOPs connect the power
narios adequately represent the actual operating environments of REGs. supply areas of different substations and provide support for intercon-
nection among different areas. Additionally, the installed capacities of
6.1.3. Seasonal typical daily scenario analysis MTs and BESs in Mode 4 which considers the optimal operation of
The initial scenarios under each season generated by CWGAN-GP are MCIESs are 8398 kW and 3379 kWh, respectively. Both values are
reduced to four typical scenarios using K-Medoids, the seasonal typical reduced compared to those in Mode 3, which are 12515 kW and 3529
scenarios of REGs and their probability are presented in Fig. 8 and kWh, respectively. The same applies to Modes 1 and 2. This is because
Table 4, respectively. MCIESs can be used as bi-directionally regulated flexible resources to
From Fig. 8, it can be observed that the PV outputs are generally provide flexibility support for the ADN, thus reducing the ADN

Fig. 9. ADN expansion planning results under different modes.

11
C. Wang et al. Energy 307 (2024) 132601

Table 5
ADN expansion planning schemes under different modes.
Mode New/expanded substations Newly built lines WT/kW (bus) PV/kW MT/kW BES/kWh/kW SOP/kVA
(MVA) (bus) (bus) (bus) (bus)
Same Different

Mode 1 S3 (22.2), 19-20, 18–19, 44–45, S4- 800(14), 5000 4700 (5), 442 (9), 63/16 (16), /
S4 (22.2) 17–18, 30, (17), 5000 (10), 1089 (13), 734/184 (32)
18-21, 22–23, S4-22, S4-21, 5000(24), 5000 5000 (33), 1628 (16),
24–25, 31–37 (39), 1100 (37) 186 (32)
8-25, 33–39, 3400 (47)
Mode 2 S3 (22.2), 32–39, 44–45, S4- 200(12), 5000 4300 (8), 442 (9), 275/69 (32) /
S4 (22.2) 38-39, 38–44, 30, (17), 5000 (10), 1075 (13),
26–27, S4-22, S4-21, 5000(24), 5000 1200 (30), 1338 (16),
27–28, S3-28, S3- 31–37 (39), 5000 (33) 301 (32)
36, 4300 (47)
Mode 3 S3 (22.2) 35–36, 34–35, S3- 2-8, 41–42, 4500(12), 2500 1300 (10), 3600 (9), 1441/360 (8), 2000 (46–47),
41, 31-37, 9–17, (14), 2200 (33), 1690 (13), 88/22 (17), 1400 (2–8),
16-40, 14–46, 23-24, 8–33, 4900(17), 5000 4600 (37) 1100 (16), 2000/500 (50) 1200 (10–15)
46–47, 13-43, 10-15 (24), 6125 (32)
42-47, 42–48, 5000(39), 5000
48–49, (47)
Mode 4 S3 (22.2) 49-50, 12–45, 9-22, 10-31 4900(12), 5000 3800 (10), 4660 (9), 2000/500 (13), 1600 (22–23),
37–43, 6-28, 9–17, (14), 900 (26), 669 (13), 1379/345 (50) 2900 (6–28),
30-43, 29-30 23-24, 8–33, 5000(17), 5000 2000 (33) 3069 (32) 1600 (10–15)
10-15, 13-43 (24),
5000(39), 3400
(47)

Table 6 Table 7
ADN expansion planning costs under different modes. Network flexibility indicators under different modes.
Annual cost (×104 $) Mode 1 Mode 2 Mode 3 Mode 4 Mode Average branch load Load uniformity Average voltage
ratio (%) (%) deviation (%)
Total cost 2505.31 2480.86 2370.72 2349.95
Total operation cost 1886.79 1862.91 1869.32 1850.22 Mode 21.21 34 4.03
Total investment cost 618.52 617.95 501.40 499.73 1
Line investment cost 18.35 18.35 21.36 22.11 Mode 23.63 30.13 4.25
Substation investment cost 348.00 348.00 174.00 174.00 2
SOP investment cost 0 0 6.79 9.01 Mode 19.19 25.18 3.51
PV investment cost 84.49 82.88 43.31 35.83 3
WT investment cost 159.18 161.67 223.02 234.63 Mode 15.47 22.64 2.72
BES investment cost 1.62 0.56 7.17 6.87 4
MT investment cost 6.88 6.49 25.75 17.28
Power purchase cost from the main 1768.83 1796.6 1492.65 1602.07
grid configuration for other flexible resources while meeting the load
Fuel cost of MT 71.10 71.00 305.54 192.06 demand.
O&M cost 45.42 45.09 69.68 63.93
Flexible resource interaction cost 0 − 51.23 0 − 8.98
with MCIES 6.3. Economic analysis
Insufficient flexibility penalty cost 1.44 1.45 1.45 1.14

The total cost and details of the ADN expansion planning under
different modes are given in Table 6.
Table 6 indicates that Mode 4 (the proposed model) exhibits the
lowest insufficient flexibility penalty cost and total cost, whereas Mode 1
(without consideration of the MCIES and SOPs) has the highest total

Fig. 10. Insufficient flexibility penalty cost and total cost under different
confidence levels.

Fig. 11. ATC iteration curve.

12
C. Wang et al. Energy 307 (2024) 132601

cost. Compared with Mode 4, Mode 3 does not consider the optimal introduces a flexibility-based expansion planning approach for ADNs,
operation of the MCIES, resulting in a decrease in the sales revenue of which effectively utilizes the operation flexibility of MCIES to postpone
the ADN to the MCIES. Additionally, due to the lack of flexibility support ADN investment and realizes the coordinated planning of DERs and
from the MCIES to ADN, the initial investment cost of flexible resources power networks, thereby improving the economy, network flexibility,
is increased. Two new substations are built in Mode 2 (without consid- and supply-demand flexibility of the overall ADN. Responding to
ering SOPs) to meet the load demand. Although this improves the renewable uncertainties, the developed model-free CWGAN-GP scenario
interaction benefits with the MCIES, the investment cost of the sub- generation method portrays the correlation, volatility, and uncertainty
stations and the power purchase cost from the main network are characteristics of REG outputs in different seasonal environments and
increased, resulting in a higher total cost. This indicates that the pro- does not require explicit feature model construction, which has broad
posed model improves the flexibility and mutual aid ability of various applicability and provides a new approach for planning scenario gen-
power supply areas in the ADN by considering SOPs and the MCIES, eration. Finally, the proposed ATC-based distributed modeling method
while ensuring economic efficiency and improving the operational refines the interest game among all stakeholders regarding economic
flexibility of the entire ADN. dispatch on the basis of ensuring privacy, truly reflecting the advantages
of decentralized autonomy of MCIESs.
6.4. Insufficient flexibility risk analysis In future studies, the reactive power optimization and regulation
strategy of MCIES should be further considered during ADN planning. In
To analyze the impact of insufficient flexibility risk on ADN expan- addition, the proposed method provides a planning strategy based on
sion planning, a sensitivity analysis on confidence level of the CVaR current system conditions. However, this method is widely applicable
model is conducted, which is shown in Fig. 10. and can be adapted to the modeling of real-world systems. To realize
The confidence level reflects the flexibility planning requirements of long-term load growth or increased renewable energy integration, the
ADNs. Fig. 10 shows that as the confidence level increases, the insuffi- proposed planning model can be re-leveraged to determine the alloca-
cient flexibility penalty cost of the ADN gradually decreases, whereas tion schemes for increased REGs and flexible resources.
the total cost gradually increases; above the 95 % confidence level, the
total cost increases significantly. This is because the increase in system CRediT authorship contribution statement
operational flexibility comes at the expense of the economy, high-
lighting the need to balance the two factors by leveraging the indicator Chunling Wang: Writing – review & editing, Writing – original
of insufficient flexibility risk. draft, Visualization, Software, Methodology, Formal analysis. Chunm-
ing Liu: Writing – review & editing, Supervision, Methodology, Formal
6.5. Network flexibility analysis analysis, Conceptualization. Xiulin Zhou: Methodology, Investigation,
Formal analysis, Conceptualization. Gaoyuan Zhang: Visualization,
In addition to supply–demand flexibility, ADNs should also have Supervision, Formal analysis, Conceptualization.
superior network flexibility to ensure the effective transmission of
power. Therefore, the average branch load ratio, load uniformity, and Declaration of competing interest
average voltage deviation indices of ADN are utilized to quantify and
evaluate the network flexibility, which is presented in Table 7. The authors declare that they have no known competing financial
From Table 7, it is evident that the average branch load ratio is the interests or personal relationships that could have appeared to influence
lowest in Mode 4, i.e., the branch load margin is the largest, indicating the work reported in this paper.
that MCIESs and SOPs can improve the load distribution of branches and
transfer capacity of lines. Compared with Mode 1, the load uniformity of Data availability
Mode 4 decreases from 34 % to 22.64 %. The smaller the load unifor-
mity, the more balanced the load ratio of each branch, thus strength- The authors do not have permission to share data.
ening the ability to cope with changes in system uncertainty. Therefore,
the ADN planning scheme in Mode 4 exhibits the best network flexibility References
among the four modes. Moreover, the average voltage deviation of Mode
4 is 2.72 %, indicating that the proposed model effectively suppresses [1] Suryakiran BV, Nizami S, Verma A, et al. A DSO-based day-ahead market
mechanism for optimal operational planning of active distribution network. Energy
the voltage fluctuation and improves the ADN voltage quality compared 2023;282:128902.
with the other three modes. [2] Huang M, Chang J, Guo A, et al. Cascade hydropower stations optimal dispatch
considering flexible margin in renewable energy power system. Energy 2023;285:
129375.
6.6. Analytical target cascading algorithm analysis [3] Wang J, Hu Z, Xie S. Expansion planning model of multi-energy system with the
integration of active distribution network. Appl Energy 2019;253:113517.
The ATC-based model solution in this study converges after 10 it- [4] Zhang D, Zhu H, Zhang H, et al. Multi-objective optimization for smart integrated
energy system considering demand responses and dynamic prices. IEEE Trans
erations. Fig. 11 shows the variation in the maximum gap between the Smart Grid 2022;13(2):1100–12.
ADN and MCIES for each iteration. It can be seen that the gap between [5] Li Y, Feng B, Wang B, et al. Joint planning of distributed generations and energy
the two continuously converges to zero during the iteration process, storage in active distribution networks: a bi-level programming approach. Energy
2022;245:123226.
which illustrates the effectiveness of the ATC algorithm. In addition, the [6] Koutsoukis N, Georgilakis P. A multistage distribution network planning method
calculation time of the example is 4.2 h, indicating that the algorithm is considering distributed generation active management and demand response. IET
applicable for practical planning problems that require long-term deci- Renew Power Gener 2022;16(1):65–76.
[7] Xu X, Niu D, Peng L, et al. Hierarchical multi-objective optimal planning model of
sion processing. Notably, considering additional CIESs in the model will
active distribution network considering distributed generation and demand-side
not significantly increase the calculation time owing to the parallel response. Sustain Energy Technol Assessments 2022;53:102438.
processing capability of the ATC method. [8] Pamshetti V, Singh S, Thakur A, et al. Cooperative operational planning model for
distributed energy resources with soft open point in active distribution network.
IEEE Trans Ind Appl 2023;59(2):2140–51.
7. Conclusion [9] Hemmati R. Dynamic expansion planning in active distribution grid integrated
with seasonally transferred battery swapping station and solar energy. Energy
As coupling among diverse energy sources intensifies, various forms 2023;277:127719.

of CIES are emerging within ADNs. Tailoring to the attributes of MCIESs

characterized by abundant flexible and controllable resources, this study

13
C. Wang et al. Energy 307 (2024) 132601

[10] Huang Z, Xu Y, Chen L, et al. Coordinated planning method considering flexible [24] Pan Z, Yu W, Yi X, et al. Recent progress on generative adversarial networks
resources of active distribution network and soft open point integrated with energy (GANs): a survey. IEEE Access 2019;7:36322–33.
storage system. IET Gener Transm Distrib 2023;17(23):5273–85. [25] Wang Z, Wang S, Zhou C, et al. AVO inversion based on closed-loop multitask
[11] Tian Z, Li X, Niu J, et al. Enhancing operation flexibility of distributed energy conditional Wasserstein generative adversarial network. IEEE Trans Geosci Rem
systems: a flexible multi-objective optimization planning method considering long- Sens 2023;61:1–13.
term and temporary objectives. Energy 2024;288:129612. [26] Gulrajani I, Ahmed F, Arjovsky M, et al. Improved training of Wasserstein GANs.
[12] Yang Z, Han G, Yang F, et al. A distribution network planning method considering NIPS; 2017. p. 30.
the distributed energy resource flexibility of virtual power plants. Sustainability [27] Badoozadeh S, Nikdel N, Galvani S, et al. Probabilistic optimal power flow in wind
2023;15(19):14399. energy integrated power system based on the K-medoids data clustering method
[13] Shirsat A, Tang W. Data-driven stochastic model predictive control for DC-coupled considering correlated uncertain variables. IET Renew Power Gener 2023;17(13):
residential PV-storage systems. IEEE Trans Energy Convers 2021;36(2):1435–48. 3179–94.
[14] Wang B, Wang X, Zhu Z, et al. Siting and sizing of energy storage for renewable [28] Zhang J, Foley A, Wang S. Optimal planning of a soft open point in a distribution
generation utilization with multi-stage dispatch under uncertainty: a tri-level network subject to typhoons. Int J Electr Power Energy Syst 2021;129:106839.
model and decomposition approach. Appl Energy 2023;344:121286. [29] Ehsanbakhsh M, Sepasian M. Simultaneous siting and sizing of soft open points and
[15] Wang Q, Zhang X, Yi C, et al. A novel shared energy storage planning method the allocation of tie switches in active distribution network considering network
considering the correlation of renewable uncertainties on the supply side. IEEE reconfiguration 2022;17(1):263–80.
Trans Sustain Energy 2022;13(4):2051–63. [30] Fan W, Tan Z, Li F, et al. A two-stage optimal scheduling model of integrated
[16] Hassaballah E, Keshta H, Abdel-Latif K, et al. A novel strategy for real-time optimal energy system based on CVaR theory implementing integrated demand response.
scheduling of grid-tied microgrid considering load management and uncertainties. Energy 2023;263:125783.
Energy 2024;299:131419. [31] Long C, Wu J, Thomas L, et al. Optimal operation of soft open points in medium
[17] Wang B, Wang X, Zhu Z, et al. Siting and sizing of energy storage for renewable voltage electrical distribution networks with distributed generation. Appl Energy
generation utilization with multi-stage dispatch under uncertainty: a tri-level 2016;184:427–37.
model and decomposition approach. Appl Energy 2023;344:121286. [32] Cao W, Wu J, Jenkins N, et al. Operating principle of soft open points for electrical
[18] Yoo J, Son Y, Yoon M, et al. A wind power scenario generation method based on distribution network operation. Appl Energy 2016;164:245–57.
Copula functions and forecast errors. Sustainability 2023;15(23):16536. [33] Liu C, Wang C, Yin Y, et al. Bi-level dispatch and control strategy based on model
[19] Dong W, Chen X, Tang Q. Data-driven scenario generation of renewable energy predictive control for community integrated energy system considering dynamic
production based on controllable generative adversarial networks with response performance. Appl Energy 2022;310:118641.
interpretability. Appl Energy 2022;308:118387. [34] Li Y, Wang B, Yang Z, et al. Optimal scheduling of integrated demand response-
[20] Yuan R, Wang B, Sun Y, et al. Conditional style-based generative adversarial enabled community-integrated energy systems in uncertain environments. IEEE
networks for renewable scenario generation. IEEE Trans Power Syst 2023;38(2): Trans Ind Appl 2022;58(2):2640–51.
1281–96. [35] Li Y, Wang C, Li G, et al. Improving operational flexibility of integrated energy
[21] Wang C, Liu C, Chen J, et al. Cooperative planning of renewable energy generation system with uncertain renewable generations considering thermal inertia of
and multi-timescale flexible resources in active distribution networks. Appl Energy buildings. Energy Convers Manag 2020;207:112526.
2024;356:122429. [36] Qiu H, Gu W, Xu Y, et al. Robustly multi-microgrid scheduling: stakeholder-
[22] Li Y, Li J, Wang Y. Privacy-preserving spatiotemporal scenario generation of parallelizing distributed optimization. IEEE Trans Sustain Energy 2020;11(2):
renewable energies: a federated deep generative learning approach. IEEE Trans Ind 988–1001.
Inf 2022;18(4):2310–20. [37] Zhang S, Wang H, Li R, et al. Active distribution network expansion planning
[23] Chen J, Yang Y, Tang X, et al. Generative adversarial networks: a literature review method considering the integration of soft open point. Proceedings of the CSEE
2020;14(12):4625–47. 2023;43(1):48–60 (in Chinese).

14