Electrical Power and Energy Systems 157 (2024) 109766

Contents lists available at ScienceDirect

International Journal of Electrical Power and Energy Systems

journal homepage: www.elsevier.com/locate/ijepes

Optimal scheduling of electric vehicle ordered charging and discharging

based on improved gravitational search and particle swarm optimization
algorithm
Kui Pan, Chuan-Dong Liang, Min Lu ∗
School of Mechano-Electronic Engineering, Shihezi University, Xinjiang, 832003, China

ARTICLE INFO ABSTRACT

Keywords: With the rapid growth of the number of Electric Vehicles (EVs), access to large-scale EVs will bring serious
Electric vehicles safety hazards to the operation planning of the power system. It needs to be supported by an effective EV
Multi-objective optimization charging and discharging behavior control strategy to meet the operation demand of the power system.
Power grid
An optimization model with the objectives of minimizing grid load variance and minimizing user charging
Particle swarm optimization
cost is established. An improved hybrid algorithm is proposed for the optimal allocation of charging and
Gravitational search algorithm
discharging power of EVs by combining particle swarm optimization (PSO) algorithm and gravitational search
algorithm (GSA). The performance of variant algorithm is tested using CEC2005 benchmarking functions sets
and applied to the solution of the ordered charge–discharge optimal scheduling model. The results show that
the convergence accuracy of the algorithm is better than the traditional algorithm, and it can effectively
balance exploration and exploitation ability of the particles. In addition, the scheduling analysis is performed
for different charging strategies of EVs. The scheduling results show that with the same optimization weights,
implementing the ordered charging and discharging strategy can significantly reduce the charging cost of users
and the load variance of the grid. Thus, the operational stability of the grid and the economic benefits for
users are improved.

1. Introduction obtain the economic benefits of participating in V2G mode through the
response tariff incentive mechanism [6]. Therefore, how to reasonably
In recent years, the rapid development of EVs has been driven by control the charging and discharging behavior of EVs to avoid negative
the cleanliness of energy sources, the efficiency of power batteries, impacts on the grid while improving the economic benefits of users
and the dramatic cost reductions and performance improvements of becomes a critical issue to be solved [7].
electric drive technologies [1]. As a new generation of transportation, The charging and discharging scheduling problem of EVs has at-
EVs have unparalleled advantages over traditional vehicles in terms tracted extensive attention from scholars. Regarding the objectives,
of energy saving and emission reduction and reducing the dependence the grid side mainly aims to improve operational stability and reduce
of the transportation industry on fossil fuels [2]. However, due to the power losses. The optimization indexes include power loss, load fluctu-
randomness and uncertainty of EV users’ charging behavior [3], large- ation variance and peak-to-valley difference. For example, literature [8]
scale EVs charging disordered can have severe negative impacts on the proposed a two-stage V2G control scheme to dynamically adjust the
power grid. Such as power quality degradation, peak load increase, charging and discharging power to cut the peak load on the grid by
excessive harmonic injection and power loss [4]. predicting the load distribution of EVs. Literature [9] used cooperative
Vehicle-to-grid(V2G) technology can realize a two-way energy ex- and non-cooperative game methods to motivate EVs to participate
change between EVs and the grid. From the grid’s perspective, EVs in grid frequency regulation. Literature [10] minimizes energy loss
can be equated as distributed energy storage units to participate in due to charging overlap by optimizing the charging and discharging
grid regulation by charging and discharging. It discharges during the periods of electric vehicles. On the user side, the driving factors that
motivate users to participate in scheduling mainly come from economic
peak load period and charges during the low load period of the power
benefits and optimization metrics such as charging cost and battery
system. Realize peak and valley reduction of the system, thus improving
loss. Literature [11] proposed a globally optimal and locally optimal
the stability of the grid operation [5]. In addition, EV users can also

∗ Corresponding author.
E-mail addresses: [email protected] (K. Pan), [email protected] (C.-D. Liang), [email protected] (M. Lu).

https://doi.org/10.1016/j.ijepes.2023.109766
Received 13 June 2023; Received in revised form 18 October 2023; Accepted 28 December 2023
Available online 20 January 2024
0142-0615/© 2024 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
K. Pan et al. International Journal of Electrical Power and Energy Systems 157 (2024) 109766

scheduling scheme for EV charging and discharging to reduce the total

charging cost. Literature [12] considers the relationship between EV
charging power and state of charge to develop an optimization model
that minimizes the cost of charging. In [13], a minimum charging
cost model including battery degradation cost and parking cost was
developed to alleviate the pressure on the grid by controlling the
charging and discharging behavior of EV users. The single-objective op-
timization model can hardly take into account the interests of the grid
side and the customer side. Some scholars have built multi-objective
optimization models to balance the interests of different stakeholders.
For example, literature [7,14] proposed an optimization model to
minimize the mean squared difference of grid load and charging cost.
Literature [15] guided the orderly charging and discharging of electric
vehicles by optimizing dynamic tariffs in order to achieve the multiple Fig. 1. The pdf of starting time of charging.
objectives of reducing total operating costs, curtailing load peaks, and
reducing voltage fluctuations. Literature [16] regulated EV charging
and discharging loads in conjunction with electricity prices to achieve
The IGSAPSO proposed in this paper aims to optimize the charg-
the dual goals of cutting charging costs and reducing the peak-to-valley
ing and discharging behavior of EVs. The scheduling results show
difference of electricity. Literature [17] develops a mixed-integer linear
that, compared with the traditional algorithm, solving the optimization
programming model for electric vehicle charge/discharge regulation
model using IGSAPSO can significantly reduce the charging cost of EV
and battery degradation to improve the economic returns of using and
users. Moreover, it effectively suppresses the load fluctuation of the grid
multi-party energy trading. In summary, the critical issues in managing and improves the safety and stability of the grid. The main contribu-
EV charging are how to improve the operational quality of the grid and tions of this paper are two. (1) An optimization model for EV ordered
reduce the cost of charging for users. charging and discharging and a scheduling strategy are proposed. This
The ordered charging and discharging scheduling of electric ve- model can simultaneously reduce the charging cost of EV users and
hicles is a nonlinear and complex non-convex optimization problem, the load fluctuation of the grid. (2) Based on the characteristics of
which is a complex optimization problem with multiple objectives and the optimization model, an effective intelligent hybrid optimization
constraints. Currently, the commonly used algorithms include genetic algorithm (IGSAPSO) is proposed. The performance of this algorithm
algorithm(GA) [18], Tabu search (TS) algorithm [19], artificial bee is tested by benchmark functions to find the optimal performance and
colony (ABC) algorithm [20], gravitational search algorithm (GSA) [21, applied to the solution of the EV charging and discharging optimization
22], particle swarm optimization (PSO) algorithm [23,24], etc. GSA is model. In addition, different charging strategies, different optimization
extremely capable of local exploitation and weak in global exploration. weights and different number of EVs responding to V2G mode for EVs
PSO has fast convergence and high search efficiency, but weak local are scheduled and analyzed.
exploration ability, which makes the particles easy to fall into local
optimum prematurely. Therefore, combining the two algorithms of GSA 2. Charging model of EVs
and PSO to form a hybrid, team learning-based parallel algorithm can
fully utilize the advantages of both algorithms. In literature [25], the For simplicity of calculation, in this paper, a day is discretized into
self-awareness capability of the PSO algorithm is partially replaced 96 time slots, and the duration of each time slot is 15 min. the EV
charging load modeling process is as follows:
by the acceleration of the GSA, but the algorithm ignores the mem-
ory function of the particles. In literature [26], simplifies the search
2.1. Electric vehicle charging model
space for the optimal parameters of the algorithm, but the algorithm
is only applicable to beam map optimization problems and is not
Research shows that over 80% of EV users are used to charging
generalizable. Literature [27] introducing acceleration of gravitational
at home at night and only charge once a day on average [32,33].
search algorithm in particle swarm velocity update strategy forming Therefore, when modeling the EV charging load, it is assumed that
a new hybrid algorithm. Literature [28] sets the inertia weights as a the time the user returns home after the last trip is when the EV is
function of the fitness function values of the global and the local best connected to the grid, and the next trip is when the EV leaves the grid.
uses an adaptive learning factor to control the particle update rate. In
literature [29,30], the acceleration part of GSA is introduced into the 2.1.1. Starting of charging time
iterative velocity update formula of PSO for superposition to form a To model the starting of charging time, assuming that the probabil-
new velocity update strategy. However, the hybrid algorithms of PSO ity density function (pdf) [24] is defined as in Eq. (1):
and GSA are based on simple teamwork, and the local development
∑
𝑆
( )2
capability of GSA and the global memory of PSO are not fully utilized. 𝑓𝑠 (𝑥𝑠 ) = 𝑎𝑠 exp[− 𝑥𝑠 − 𝑏𝑠 ∕𝑐𝑠 ] (1)
Literature [31] retains the global memory function of particles and 𝑠=1
introduces the acceleration of GSA into the individual cognitive part where a, b, and c are the peak, peak position, and half-width informa-
of PSO. However, the algorithm reduces the gravitational interaction tion of the Gaussian curve, respectively, the specific parameters are the
between particles and weakens the local exploitation ability of the same as in the literature [24]. Fig. 1 shows the pdf of starting time of
algorithm. Based on this, an improved GSAPSO (IGSAPSO) is proposed charging.
to replace the acceleration part of GSA with the individual cognitive
part of PSO. The individual and social cognitive parts are also combined 2.1.2. Starting of charging time
to form a new speed update strategy. The elite learning strategy is in- To model the ending of charging time, assuming the probability
troduced to enhance the global exploration capability of the algorithm. density function [24] is defined as in Eq. (2):
The exploration and exploitation ability of the algorithm is balanced by ∑
𝐸
( )2
adaptive inertia weights and learning factors at different stages of the 𝑓𝑒 (𝑥𝑒 ) = 𝑎𝑒 exp[− 𝑥𝑒 − 𝑏𝑒 ∕𝑐𝑒 ] (2)
optimization search process. The new hybrid algorithm can fully utilize 𝑒=1

the advantages of PSO and GSA. Fig. 2 shows the pdf of ending time of charging.

2
K. Pan et al. International Journal of Electrical Power and Energy Systems 157 (2024) 109766

Fig. 4. Schematic diagram for determining EV participation scheduling conditions.

during disordered charging. ⌈𝑥⌉ is the maximum integer not greater

than x.
The relationship between the dwell time of an EV and the actual
Fig. 2. The pdf of ending time of charging. time required for charging is the determining factor in determining
whether the EV has the conditions to participate in ordered charging
and discharging schedule. The relationship is shown in Fig. 4. If 𝑇𝑚𝑐 ≤
𝑇𝑚𝑝 , the vehicle does not meet the conditions to participate in charging
and discharging schedule, and the system will schedule it for disordered
charging. If 𝑇𝑚𝑐 > 𝑇𝑚𝑝 , the scheduling system develops the charging and
discharging scheme for this EV, and the user determines whether to
participate in the scheduling.

2.2. Rolling optimization strategy for ordered charging and discharging of

Fig. 3. Two cases of EV dwell duration. EVs

Assuming that the total number of EVs available for dispatch in the
2.1.3. Dwell duration region is M. Whenever a new electric vehicle (𝐸𝑉𝑚 ) is connected to the
There are two cases of EV users accessing and leaving the grid, as grid, the scheduling system follows the following steps to achieve the
shown in Fig. 3. The first refers to the EV accessing/leaving the grid on ordered charging and discharging of 𝐸𝑉𝑚 .
the same day. The other one is accessing the grid on the first day and
leaving the grid on the second day. The dwell period 𝑇𝑚𝑃 is calculated 2.2.1. Obtaining 𝐸𝑉𝑚 charging requirements
by Eq. (3). The scheduling system obtains the battery capacity 𝐶𝑚 of the 𝐸𝑉𝑚
{ ⌊ ⌋ and the current 𝑆𝑂𝐶𝑚𝑎𝑐 of the battery through the EV’s battery manage-
𝑝
𝑇𝑚𝑑𝑒 − 𝑇𝑚𝑎𝑐 , 1 ≤ 𝑇𝑚𝑎𝑐 < 𝑇𝑚𝑑𝑒 ≤ 96 ment system. Record the time 𝑇𝑚𝑎𝑐 of the vehicle was connected to the
𝑇𝑚 = ⌊ ⌋ (3)
96 + 𝑇𝑚𝑑𝑒 − 𝑇𝑚𝑎𝑐 , 1 ≤ 𝑇𝑚𝑑𝑒 < 𝑇𝑚𝑎𝑐 ≤ 96 grid. In order to make a reasonable charging and discharging plan for
the user, the user needs to enter the time 𝑇𝑚𝑑𝑒 of leaving the grid and
where 𝑇𝑚𝑎𝑐 , 𝑇𝑚𝑑𝑒 is the time of accessing and departing the grid, respec- the expected charge state 𝑆𝑂𝐶𝑚𝑑𝑒 during the period of leaving in the
tively; ⌊𝑥⌋ is the smallest integer not less than x. system.

2.1.4. Charge/discharge control factor 2.2.2. Develop 𝐸𝑉𝑚 charging and discharging schedules based on customer
The control coefficients are introduced to facilitate the control of demand and the operational status of the distribution network
the charging and discharging behavior of the EV: The scheduling system calculates the actual period 𝑇𝑚𝑐 required to
complete charging of 𝐸𝑉𝑚 based on the status information of 𝐸𝑉𝑚 and
𝑋𝑚 = 𝑧𝑒𝑟𝑜𝑠[𝑥1,𝑚 , 𝑥2,𝑚 , … , 𝑥96,𝑚 ] (4) charging demand. When the conditions for participating in charging
where 𝑋𝑚 is the set that controls the charging and discharging behavior and discharging schedule are met, the system calculates the charging
of the vehicle EVm different moments. zeros denotes that 𝑋𝑚 is ini- and discharging power of 𝐸𝑉𝑚 for each period and the total charging
tialized as a zero vector, 𝑥𝑚,1 ∼ 𝑥𝑚,96 is the charging and discharging cost of users based on the current load state of the grid, and users
control coefficients for each time period of this EV to control the choose independently whether to respond to the V2G scheduling plan.
charging and discharging behavior and charging and discharging power If the user does not meet the conditions for participation in charging
and discharging scheduling or refuses to participate, the system will
of the vehicle at each moment. The control rule is defined as Eq. (5).
arrange for disordered charging.
⎧0<𝑥 ≤1 𝑐ℎ𝑎𝑟𝑔𝑖𝑛𝑔
⎪ 𝑡,𝑚
2.2.3. Rolling optimization of EVs
⎨ 𝑥𝑡,𝑚 = 0 𝑠𝑡𝑎𝑡𝑖𝑜𝑛𝑎𝑟𝑦 (5)
⎪−1 ≤ 𝑥𝑡,𝑚 < 0 𝑑𝑖𝑠𝑐ℎ𝑎𝑟𝑔𝑖𝑛𝑔 The charging and discharging guiding strategy proposed in this pa-
⎩ per is rolling over time. That is, the charging and discharging schedule
is developed in real-time according to the state of EVs connected to
2.1.5. Actual time required for charging
the grid at each moment to smooth out the load fluctuation of the
When 𝐸𝑉𝑚 is connected to the grid, the scheduling system needs
grid. The key is that the system updates the grid’s load status at the
to calculate the actual period 𝑇𝑚𝑐 required for charging at maximum
beginning of each moment. When a new 𝐸𝑉𝑚 is detected, the system
power (also the period required for disordered charging) based on the
makes a charging and discharging plan for the next 𝑇𝑚𝑝 period based
charging demand of the user and the current state of the vehicle, as on the 𝐸𝑉𝑚 status information and the grid condition. The system will
shown in Eq. (6). strictly implement the charging and discharging plan after the user
⌈( ) ( ) ⌉ confirms it. So on and so forth, the charging and discharging power
𝑇𝑚𝑐 = (𝑆𝑂𝐶𝑚𝑑𝑒 − 𝑆𝑂𝐶𝑚𝑎𝑐 ) ⋅ 𝐶𝑚 ∕ 𝜂𝑃𝑐,max × 4 (6)
of EVs connected to the grid at different periods is regulated according
where 𝑆𝑂𝐶𝑚𝑎𝑐 , 𝑆𝑂𝐶𝑚𝑑𝑒 is the state of charge at the time of access- to the latest grid load condition. To achieve the smoothing of the grid
ing/leaving the grid, respectively. 𝜂 is the charging efficiency. 𝑃𝑐,𝑚𝑎𝑥 load curve and reduce charging cost for users. The specific process is
is the maximum charging power of the 𝐸𝑉𝑚 , which is also the power shown in Fig. 5.

3
K. Pan et al. International Journal of Electrical Power and Energy Systems 157 (2024) 109766

Fig. 5. Electric vehicle charging and discharging rolling optimization flow chart.

3. Optimization scheduling model for EVs 𝐸𝑉 is the

of the grid at the period t. M is the total number of EVs. 𝑃𝑡,𝑚
charging and discharging power of 𝐸𝑉𝑚 at the period t, and 𝑃𝑇𝑎𝑣𝑔 is the
The charging and discharging scheduling of electric vehicles is a average load of the grid in a day.
multidimensional, multivariate, and multi-constrained nonlinear opti-
mization problem. In this chapter, the stable operation of the grid and 3.1.2. Economic costs for electric vehicle users
the actual interests of users are fully considered according to the actual Cyclic charging and discharging of EVs can bring about battery per-
situation. Two sub-functions are selected to form a multi-objective formance degradation problems. Therefore, the cost of battery degrada-
function, which are the minimum variance of load fluctuation in the tion is an essential factor affecting the willingness of users to participate
distribution network and the minimum economic cost for EV users. The in ordered charge/discharge scheduling. Literature [34] gives a plot of
constraints of the two different objectives and models are specified as the number of battery cycles versus lifetime (values given by the battery
follows: manufacturer) and the cost of battery degradation as shown in Eq. (8).
{ 𝑉 2𝐺 = |𝐵 ∕100 | ⋅ 𝛤 ∕𝐶 ⋅ 𝐶 𝐵
𝐶𝑡,𝑚 | 𝑚 | 𝑡,𝑚 𝑚
3.1. Objective function { } (8)
𝛤𝑡,𝑚 = max 0, (𝑆𝑂𝐶𝑡−1,𝑚 − 𝑆𝑂𝐶𝑡,𝑚 ) ⋅ 𝐶𝑚
3.1.1. Load variance of the grid where 𝐶𝑡,𝑚𝑉 2𝐺 is the cost (¥) of battery degradation of the vehicle 𝐸𝑉
𝑚
Smoothing the load fluctuation of the grid can vastly reduce the at the period t. 𝐵𝑚 is the coefficient of the linear relationship between
system losses and improve the operational stability of the grid. The load battery life and the number of cycles, i.e., the slope of the fitted curve
variance 𝑓𝑚,1 of the grid after the 𝐸𝑉𝑚 participates in the dispatch is in Fig. 6. 𝛤𝑡,𝑚 is the cycle charge and discharge power (kW h), 𝐶 𝐵 is
defined by Eq. (7). the battery replacement cost (¥). 𝑆𝑂𝐶𝑡−1,𝑚 , 𝑆𝑂𝐶𝑡,𝑚 is the charge state
( )2 of the vehicle 𝐸𝑉𝑚 at the period t – 1 and at the period t, respectively.
⎧ ∑𝑇 ∑𝑀
⎪min 𝑓𝑚,1 = 𝑃𝑡𝐹 𝑑 + 𝐸𝑉
𝑃𝑡,𝑚 − 𝑃𝑇𝑎𝑣𝑔 𝐶𝑚 is the 𝐸𝑉𝑚 battery capacity.
⎪ 𝑡=1 𝑚=1 Considering the cost of battery decline, the second objective func-
⎨ ( ) (7)
⎪ ∑𝑇 ∑𝑀 tion is the lowest charging cost 𝑓𝑚,2 of 𝐸𝑉𝑚 , it is defined as in Eq. (9):
𝑎𝑣𝑔 𝐹𝑑 𝐸𝑉
⎪ 𝑃 = 1∕𝑇 𝑃 + 𝑃 𝑇 ((
⎩ 𝑇
𝑡=1
𝑡
𝑚=1
𝑡,𝑚
⎧ ∑ ) )
𝐸𝑉 𝑉 2𝐺
⎪min 𝑓𝑚,2 = 𝐸𝑡 ⋅ 𝑃𝑡,𝑚 ∕4 + 𝐶𝑡,𝑚
where T is the number of periods of a day (the paper takes 15 min as a ⎨ 𝑡=1 (9)
⎪ { }
period and discrete the day into 96 time periods). 𝑃𝑡𝐹 𝑑 is the base load ⎩ 𝐸𝑡 ∈ 𝐸𝑐𝑝 , 𝐸𝑐𝑢 , 𝐸𝑐𝑣 , 𝐸𝑑𝑝 , 𝐸𝑑𝑢, 𝐸𝑑𝑣

4
K. Pan et al. International Journal of Electrical Power and Energy Systems 157 (2024) 109766

3.2.4. Feeder capacity constraint

The maximum allowable load exists on the feeder of the distribution
grid. When EVs are connected to the grid, the sum of the regional base
load and the charging load of all EVs connected to the grid during that
period must not exceed the maximum load capacity of the feeder.
∑
𝑀
𝑃𝑡𝐹 𝑑 + 𝐸𝑉
𝑃𝑡,𝑚 ≤ 𝑃𝑙,max (15)
𝑚=1
where 𝑃𝑙,𝑚𝑎𝑥 indicates the maximum load capacity supported by the
feeder l.

3.2.5. Schedulable time constraint

Fig. 6. Cycle times-life curve relationship of nanometer lithium iron phosphate battery.
EVm can only be dispatched during the time it is connected to the
grid:

where 𝐸𝑡 is the charging and discharging tariff (¥/(kW h)) at the period 𝑇𝑚𝑎𝑐 ≤ 𝑇𝑉 2𝐺,𝑚 ≤ 𝑇𝑚𝑑𝑒 (16)
t ; 𝐸𝑐𝑝 , 𝐸𝑐𝑓 , 𝐸𝑐𝑣 is the charging tariff for the peak, flat and valley hours, {
∁𝑋𝑚 𝐴𝑚 = 0
respectively. And 𝐸𝑑𝑝 , 𝐸𝑑𝑓 , 𝐸𝑑𝑣 is the discharging tariff for the peak, (17)
𝐴𝑚 = {𝑥𝑚,𝑇𝑚𝑎𝑐 , 𝑥𝑚,𝑇𝑚𝑎𝑐 +1 , … , 𝑥𝑚,𝑇 𝑑𝑒 −1 }
flat and valley hours. 𝑚

where 𝑇𝑉 2𝐺,𝑚 is the set of times when the vehicle EVm participates
3.1.3. The objective function of the model in charge/discharge scheduling, and the elements in the set 𝐴𝑚 are
the charge/discharge control coefficients of the vehicle EVm at each
Based on the linear weighted sum method, the objective function
moment during the grid connection time. Eq. (17) shows that the
𝑓𝑚,1 , 𝑓𝑚,2 is normalized by the Eq. (10):
EVm’s charge/discharge control coefficients are 0 and have no charge/
{ max ) + 𝛽(𝑓
min 𝑓𝑚 = 𝛼(𝑓𝑚,1 ∕𝑓𝑚,1 max
𝑚,2 ∕𝑓𝑚,2 )
discharge behavior both before connecting to the grid and after leaving
(10) the grid.
𝛼 =1−𝛽
where 𝑓𝑚 is the multi-objective optimization function of vehicle 𝐸𝑉𝑚 . 4. Particle swarm algorithm and gravitational search algorithm
max , 𝑓 max are the maximum value of the objective function, and 𝛼, 𝛽
𝑓𝑚,1 𝑚,2
4.1. Standard particle swarm optimization algorithm
are the optimization weight.
The PSO mainly simulates the flight-foraging behavior of a flock of
3.2. Constraints birds. The core idea is to iteratively update the velocity and position
of the particles to find the optimal position in the solution space.
3.2.1. Charge and discharge power constraints The
{ position and } velocity
{ vectors of the } particles are denoted as 𝑥𝑖 =
Excessive charge/discharge power of EVs will accelerate the aging 𝑥𝑖1 , 𝑥𝑖2 , … , 𝑥𝑖𝐷 , 𝑣𝑖 = 𝑣𝑖1 , 𝑣𝑖2 , … , 𝑣𝑖𝐷 , where 𝑖 = 1, 2, … , 𝑛,n denotes
process of the battery. In order to protect the battery, the charge/ the number of particles and D denotes the dimension of the solution
discharge power of 𝐸𝑉𝑚 should be limited. space. The particle moves in the search space, constantly updating
{ its position guided by the particle inertia, the optimal solution pbest
𝐸𝑉 ≤ 𝑃 𝐸𝑉
0 ≤ 𝑃𝑐,𝑡,𝑚 𝑐 max,𝑚 searched by the individual and the optimal solution gbest of the popu-
(11)
𝐸𝑉 𝐸𝑉
0 ≤ 𝑃𝑑𝑐,𝑡,𝑚 ≤ 𝑃𝑑𝑐 lation. The iterative update of the position and velocity of the particle
max,𝑚
is defined as shown in Eq. (18).
𝐸𝑉 , 𝑃 𝐸𝑉
where 𝑃𝑐,𝑡,𝑚 𝑑𝑐,𝑡,𝑚
are the charging and discharging power of the {
𝐸𝑉𝑚 at the period t, respectively. 𝑃𝑐𝐸𝑉 𝐸𝑉 𝑣𝑑𝑖 (𝑘 + 1) = 𝑤𝑣𝑑𝑖 (𝑘) + 𝑐1 𝑟1 (𝑝𝑏𝑒𝑠𝑡𝑑𝑖 (𝑘) − 𝑥𝑑𝑖 (𝑘)) + 𝑐2 𝑟2 (𝑔𝑏𝑒𝑠𝑡(𝑘) − 𝑥𝑑𝑖 (𝑘))
max,𝑚 , 𝑃𝑑𝑐 max,𝑚 are the maximum
charging and discharging power. 𝑥𝑑𝑖 (𝑘 + 1) = 𝑥𝑑𝑖 (𝑘) + 𝑣𝑑𝑖 (𝑘 + 1)
(18)
3.2.2. EV battery capacity constraints
where w is the inertia weight coefficient (non-negative) of the particle;
The charging and discharging depth of 𝐸𝑉𝑚 also affects the life
𝑐1 , 𝑐2 are the individual learning factor and social learning factor,
of the battery, and to avoid over-charging and over-discharging the
respectively; 𝑟1 , 𝑟2 are random numbers distributed between [0,1]; k
battery, the charge state of the battery needs to be limited. is the number of iterations.
𝑆𝑂𝐶min ≤ 𝑆𝑂𝐶𝑚 ≤ 𝑆𝑂𝐶max (12) The standard PSO cannot balance the ability to explore globally
with the ability to exploit locally, making it highly susceptible to
where 𝑆𝑂𝐶𝑚𝑎𝑥 , 𝑆𝑂𝐶𝑚𝑖𝑛 is the upper and lower limit of the 𝐸𝑉𝑚 charge falling into a local optimum. Therefore, it is necessary to improve it
state. to overcome its shortcomings.

4.2. Standard gravitational search algorithm

3.2.3. User travel demand constraints
When the 𝐸𝑉𝑚 leaves the grid, the battery charge state should be no GSA is an intelligent optimization algorithm inspired by Newton’s
less than the user’s desired value to ensure that the 𝐸𝑉𝑚 power meets law of gravity in physics [35]. In the gravitational algorithm, the
the user’s travel needs. position of each celestial body represents a feasible solution, while the
mass of the celestial body reflects the goodness of that position, and
𝑆𝑂𝐶𝑚𝑒 ≤ 𝑆𝑂𝐶𝑚𝑑𝑒 ≤ 1 (13)
the better the position, the greater the mass of the celestial body. The
𝑑𝑒
celestial bodies share information through the gravitational interaction
𝑇𝑚
∑ and guide the population to the region of the optimal solution (the
𝑆𝑂𝐶𝑚𝑑𝑒 =𝜂⋅ 𝐸𝑉
𝑃𝑡,𝑚 ∕(4 ⋅ 𝐶𝑚 ) + 𝑆𝑂𝐶𝑚𝑎𝑐 (14) celestial body with the largest mass), so that the optimal solution of the
𝑡=𝑇𝑚𝑎𝑐
optimization problem can be found. During the kth iteration, according
where 𝑆𝑂𝐶𝑚𝑒 is the expected value of the 𝐸𝑉𝑚 user. to the fitness value, the inertial mass of the ith celestial body is defined

5
K. Pan et al. International Journal of Electrical Power and Energy Systems 157 (2024) 109766

by Eq. (19).
⎧ ∑𝑛

⎪ 𝑀𝑖 (𝑘) = 𝑚𝑖 (𝑘)∕ 𝑚𝑗 (𝑘)

⎨ 𝑗=1 (19)
⎪ ( ) ( )
⎩𝑚𝑖 (𝑘) = 𝑓𝑖 (𝑘) − 𝑓𝑤𝑜𝑟𝑠𝑡 (𝑘) ∕ 𝑓𝑏𝑒𝑠𝑡 (𝑘) − 𝑓𝑤𝑜𝑟𝑠𝑡 (𝑘)
where 𝑓𝑖 (k) is the fitness value of the ith celestial body, 𝑓𝑏𝑒𝑠𝑡 (𝑘), 𝑓𝑤𝑜𝑟𝑠𝑡 (𝑘)
are the optimal and worst adaptation values at the kth iteration,
respectively.
The force on the ith celestial body in d dimensional space by
celestial body j is defined as in Eq. (20).
( ( )) ( )
𝐹𝑖𝑗𝑑 (𝑘) = 𝐺(𝑘) ⋅ 𝑀𝑖 (𝑘) ⋅ 𝑀𝑗 (𝑘)∕ 𝑅𝑖𝑗 (𝑘) + 𝜀 ∕ 𝑥𝑑𝑗 (𝑘) − 𝑥𝑑𝑖 (𝑘) (20)
‖ ‖
aotic variable in where 𝑅𝑖𝑗 (𝑘) = ‖𝑥𝑗 (𝑘), 𝑥𝑖 (𝑘)‖ is the Euclidean distance
‖ ‖
between the celestial body i and the celestial body j. 𝜀 is a tiny constant,
and G(k) is the gravitational constant at the kth iteration, defined as in
Eq. (21): Fig. 7. Particle position and velocity updating of the IGSAPSO.
[ ]
𝐺(𝑘) = 𝐺0 exp −𝜓 ⋅ (𝑘∕𝐾) (21)
strategy of the algorithm is defined as in Eq. (28):
where 𝐺0 and 𝜓 are constants and K is the total number of iterations, ( )
then the combined force on the d dimensional space for the ith celestial 𝑣𝑑𝑖 (𝑘 + 1) = 𝜔(𝑘)𝑣𝑑𝑖 (𝑘) + 𝑐1 (𝑘)𝑟1 ( 𝑃𝑖𝑑 (𝑘) + 𝑎𝑑𝑖 (𝑘) ∕2 − 𝑥𝑑𝑖 (𝑘))
body is defined as in Eq. (22): + 𝑐2 (𝑘)𝑟2 (𝑃𝑔𝑑 (𝑘) − 𝑥𝑑𝑖 (𝑘)) (28)
∑
𝑛
In [36], acceleration is introduced into the social cognition part of
𝐹𝑖𝑑 (𝑘) = 𝑟𝑎𝑛𝑑(0, 1)𝑗 ⋅ 𝐹𝑖𝑗𝑑 (𝑘) (22)
𝑗=1,𝑗≠𝑖 the particle, and the inertia weights are set in an exponentially decreas-
ing way to balance the particle’s global exploration and local exploita-
According to Newton’s second law, the acceleration of the ith celes-
tion capabilities. The velocity update strategy is defined as in Eq. (29):
tial body in d dimensions is defined as in Eq. (23):
𝑣𝑑𝑖 (𝑘 + 1) = 𝑤(𝑘)𝑣𝑑𝑖 (𝑘) + 𝑐1 (𝑘)𝑟1 (𝑃𝑖𝑑 (𝑘) − 𝑥𝑑𝑖 (𝑘))
𝑎𝑑𝑖 (𝑘) = 𝐹𝑖𝑑 (𝑘)∕𝑀𝑖𝑑 (𝑘) (23)
+ 𝑐2 (𝑘)𝑟2 ((𝑃𝑔𝑑 (𝑘) + 𝑎𝑑𝑖 (𝑘))∕2 − 𝑥𝑑𝑖 (𝑘)) (29)
In the GSA, the velocity and position of object i are updated when
iterating as shown in Eq. (24). 4.4. Improved GSA and PSO hybrid algorithm
{
𝑣𝑑𝑖 (𝑘 + 1) = 𝑟𝑎𝑛𝑑(0, 1)𝑗 ⋅ 𝑣𝑑𝑖 (𝑘) + 𝑎𝑑𝑖 (𝑘)
(24) 4.4.1. Particle position and velocity update strategy
𝑥𝑑𝑖 (𝑘 + 1) = 𝑥𝑑𝑖 (𝑘) + 𝑣𝑑𝑖 (𝑘 + 1) For IGSAPSO, the update strategy for particle position and velocity
The standard GSA converges extremely fast, resulting in a weak is defined as shown in Eq. (30).
global but robust local search. According to this feature, it is combined ⎧ 𝑣𝑑𝑖 (𝑘 + 1) = 𝑤(𝑘)𝑣𝑑𝑖 (𝑘) + 𝑐1 (𝑘)𝑟1 𝑎𝑑𝑖 (𝑘)
⎪ (( ) )
with the PSO, and the gravitational acceleration mechanism in the GSA 𝑑 𝑑 𝑑
⎨+𝑐2 (𝑘)𝑟2 𝑃𝑖 (𝑘) + 𝑃𝑔 ∕2 − 𝑥𝑖 (𝑘) (30)
is introduced to improve the optimization performance of the PSO. ⎪
⎩ 𝑥𝑑𝑖 (𝑘 + 1) = 𝑥𝑑𝑖 (𝑘) + 𝑣𝑑𝑖 (𝑘 + 1)
4.3. Hybrid algorithm for PSO and GSA where 𝑐1 (k), 𝑐2 (k) denote the adaptive learning factor of the ith particle
at the kth iteration.
The hybrid algorithm of GSA and PSO is mainly composed by intro- As shown in Eq. (30), the velocity vector update strategy for par-
ducing the core mechanism of the existence of gravitational interaction ticles consists of four components. The acceleration mechanism in the
between celestial bodies in GSA based on the optimization structure GSA replaces the individual cognitive part of the PSO, and the indi-
of PSO. The two algorithms are computed in parallel in a teamwork vidual cognitive part of the particle is fused with the social cognitive
manner. part to preserve its global memory function. The effect of each part on
In [25], the individual cognitive part of the PSO velocity update the particle velocity is illustrated in Fig. 7. By adjusting the values of
equation is replaced by the acceleration part of the GSA, and the update the control parameters, the respective advantages of PSO and GSA can
equation of the PSOGSA hybrid algorithm is defined as in Eq. (25): be exploited to improve the overall optimization performance of the
hybrid algorithm.
𝑣𝑑𝑖 (𝑘 + 1) = 𝑤𝑣𝑑𝑖 (𝑘) + 𝑐1 𝑟1 𝑎𝑑𝑖 (𝑘) + 𝑐2 𝑟2 (𝑃𝑔𝑑 − 𝑥𝑑𝑖 (𝑘)) (25)
4.4.2. Chaos initialization population
In [27], a new gravitational search particle swarm hybrid algorithm
The quality of the initial particle population greatly impacts the con-
is proposed for optimizing spiral bevel gear designs. The velocity
vergence speed and accuracy of the algorithm. Using a chaotic mapping
update strategy is defined as in Eq. (26):
to initialize the population can significantly improve the optimization
𝑣𝑑𝑖 (𝑘 + 1) = 𝑤𝑣𝑑𝑖 (𝑘) + 𝑐1 𝑎𝑑𝑖 (𝑘) + (𝑐1 ∕𝑐2 )(𝑃𝑔𝑑 − 𝑥𝑑𝑖 (𝑘)) (26) performance of the algorithm compared to a pseudo-random number
generated population [37]. In this paper, we choose a simple and
In [28], the Hybrid Gravitational Search Particle Swarm Optimiza- efficient logistic mapping to initialize the particle population, which
tion algorithm (HGSPSO) utilizes adaptive acceleration coefficients to is defined as shown in Eq. (31).
control the flight of particles. The particle velocity update strategy is
defined as in Eq. (27): 𝑋𝑛+1 = 𝜇𝑋𝑛 (1 − 𝑋𝑛 ) (31)

𝑣𝑑𝑖 (𝑘 + 1) = 𝑟1 𝑣𝑑𝑖 (𝑘) + 𝑐1 𝑟2 (𝑃𝑖𝑑 (𝑘) − 𝑥𝑑𝑖 (𝑘)) + 𝑐2 𝑟3 (𝑃𝑔𝑑 − 𝑥𝑑𝑖 (𝑘)) + 𝑎𝑑𝑖 (𝑘) (27) where 𝑋𝑛 is the nth chaotic variable in the interval (0,1), 𝜇 represents
the control parameter in the range [0,4]. Fig. 8 illustrates the bifur-
In [31], good optimization was achieved by fusing the personal cation of the Logistic chaotic mapping with 𝜇 in the interval [3,4]. It
cognition part of PSO with the acceleration part of GSA. The update can be seen that the mapping state of the Logistic is closely related to

6
K. Pan et al. International Journal of Electrical Power and Energy Systems 157 (2024) 109766

𝑐𝑖 = 𝑐𝑖𝑠 + (𝑐𝑖𝑒 − 𝑐𝑖𝑠 ) ⋅ (𝑘)2 ∕(𝐾)2 , 𝑖 = 1, 2 (33)

where 𝑤𝑚𝑎𝑥 , 𝑤𝑚𝑖𝑛 are the maximum and minimum values of inertia
weights, respectively. 𝑓𝑎𝑣𝑔 is the average fitness value at the kth iter-
ation of the particle population. 𝑐𝑖𝑠 , 𝑐𝑖𝑒 are the initial and termination
values of 𝑐𝑖 .

4.4.4. Elite learning strategies

Elite learning strategy focuses on generating a new population of
merit-seeking by ranking the current set of particle solutions in terms
of the magnitude of their fitness values, generating 10% new solutions
from the 10% solutions with the best fitness values in the set, and
replacing the 10% new solutions with the 10% solutions with the worst
Fig. 8. Logistic mapping bifurcation graph with 3 ≤ 𝜇 ≤ 4. fitness in the set, keeping the intermediate solutions unchanged. The
operator for generating new solutions from the elite solutions is defined
as shown in Eq. (34).
the control parameter 𝜇. When 𝜇 = 4 the chaotic system is in a full
mapping state. ⎧ 𝑥𝑖,𝑛𝑒𝑤 = 𝑥𝑖 × 𝑄
⎪
⎨ 𝑄 = 𝑅 𝑖,𝑠𝑡𝑎𝑟 × 𝑟𝑎𝑛𝑑(−0.5, 0.5)∕𝑑 (34)
4.4.3. Adaptive inertia weights and learning factors ⎪
⎩ 𝑥𝑖,𝑤𝑜𝑟𝑠𝑡 = 𝑥𝑖,𝑛𝑒𝑤
Both the inertia weight coefficient and the learning factor can
where 𝑅𝑖,𝑠𝑡𝑎𝑟 is the Euclidean distance between the optimal solution and
balance the global exploration and local exploitation of the particle.
the closest solution, rand(−0.5, 0.5) is a random number in the range
IGSAPSO proposed in this paper adopts adaptive inertia weights and
of (−0.5, 0.5). Q is the variation factor to generate the new solution,
asymmetric learning factors to balance the relationship between par-
and 𝑥𝑖,𝑛𝑒𝑤 , 𝑥𝑖,𝑤𝑜𝑟𝑠𝑡 are the new solution and the replaced poor solution,
ticle exploration and development ability. The specific definition is
respectively.
shown in Eqs. (32)–(33).
⎧𝑤 ( ) ( )
− 𝑓𝑖 (𝑘) − 𝑓min ∕ 𝑓𝑎𝑣𝑔 − 𝑓min ⋅ (𝑤max − 𝑤min ), 4.4.5. Algorithm solving steps
⎪ max
𝑤(𝑘) = ⎨ 𝑓𝑖 (𝑘) ≤ 𝑓𝑎𝑣𝑔 (32) The flow chart of IGSAPSO algorithm is shown in Fig. 9, and the
⎪ 𝑤max , 𝑒𝑙𝑠𝑒 specific steps of the algorithm are as follows:
⎩

Fig. 9. Flowchart of the proposed IGSAPSO algorithm.

7
K. Pan et al. International Journal of Electrical Power and Energy Systems 157 (2024) 109766

Table 1
CEC-2005 benchmark functions.
Function Dim Range Minima
Unimodal benchmark function
∑𝑛
𝐹 − 1(𝑥) = 𝑖=1 𝑥2𝑖 50 [−100, 100] 0
∑𝑛 | | ∏𝑛 | |
𝐹 − 2(𝑥) = 𝑖=1 |𝑥𝑖 | + 𝑖=1 |𝑥𝑖 | 50 [−10, 10] 0
∑𝑛 (∑𝑖 )2
𝐹 − 3(𝑥) = 𝑖=1 𝑗−1 𝑗
𝑥 50 [−100, 100] 0
{ }
𝐹 − 4(𝑥) = max𝑖 ||𝑥𝑖 || 1≤𝑖≤𝑛 50 [−100, 100] 0
∑𝑛−1 [ ( )2 ( )2 ]
𝐹 − 5(𝑥) = 𝑖=1 100 𝑥𝑖+1 − 𝑥2𝑖 + 𝑥𝑖 − 1 50 [−30, 30] 0
∑𝑛−1 ([ ])2
𝐹 − 6(𝑥) = 𝑖=1 𝑥𝑖 + 0.5 50 [−100, 100] 0
∑𝑛−1
𝐹 − 7(𝑥) = 𝑖=1 𝑖𝑥4𝑖 + random [0, 1] 50 [−1.28, 1.28] 0
Multimodal benchmark function
(√ )
∑𝑛 |𝑥𝑖 |
𝐹 − 8(𝑥) = 𝑖=1 −𝑥𝑖 sin | | 50 [−500, 500] −418.9829 × 𝐷
∑𝑛 [ ( ) ]
𝐹 − 9(𝑥) = 𝑖=1 𝑥2𝑖 − ( 10 cos 2𝜋𝑥 + 10 ) 50 [−5.12, 5.12] 0
√ ∑𝑖 ( ∑ ( ))
𝑛 𝑛
𝐹 − 10(𝑥) = −20 exp −0.2 1𝑛 𝑖=1 𝑥2𝑖 − exp 1𝑛 𝑖=1 cos 2𝜋𝑥𝑖 + 20 + 𝑒 50 [−32, 32] 0
( )
1 ∑𝑛 ∏𝑛 𝑥
𝐹 − 11(𝑥) = 4000 𝑖=1 𝑖
𝑥2 − 𝑖=1 cos √𝑖 + 1 50 [−600, 600] 0
{ 𝑖 }
𝜋 ( ) ∑
𝑛−1
( )2 [ ( )] ( )2 ∑𝑛
( )
𝐹 − 12(𝑥) = 10 sin 𝜋𝑦1 + 𝑦𝑖 − 1 1 + 10 sin2 𝜋𝑦𝑖+1 + 𝑦𝑛 − 1 + 𝑢 𝑥𝑖 , 10, 100, 4
𝑛 𝑖=1 𝑖=1
𝑥𝑖+1
𝑦𝑖 = 1 +
4 50 [−50, 50] 0
⎧ 𝑘 (𝑥 − 𝑎)𝑚 𝑥𝑖 > 𝑎
( ) ⎪ 𝑖
𝑢 𝑥𝑖 , 𝑎, 𝑘, 𝑚 = ⎨ 0 −𝑎 < 𝑥𝑖 < 𝑎
( )𝑚
⎪𝑘 −𝑥 − 𝑎 𝑥𝑖 < −𝑎
⎩ 𝑖
{ ( ) ∑𝑛 ( )2 [ ( )] ( )2 [ ( )]}
𝐹 − 13(𝑥) = 0.1 sin2 3𝜋𝑥1 + 𝑖=1 𝑥𝑖 − 1 1 + sin2 3𝜋𝑥𝑖 + 1 + 𝑥𝑛 − 1 1 + sin2 2𝜋𝑥𝑛
∑𝑛 ( )
+ 𝑖=1 𝑢 𝑥𝑖 , 5, 100, 4 50 [−50, 50] 0
Fixed-dimension multimodal benchmark function
( )−1
1 ∑25
𝐹 − 14(𝑥) = 500 + 𝑗=1 ∑2 1 6 2 [−65, 65] 1
𝑗+ 𝑖=1 (𝑥𝑖 −𝑎𝑖𝑗 )

∑11 [ ]
𝑥 (𝑏2 +𝑏𝑖 𝑥2 ) 2
𝐹 − 15(𝑥) = 𝑗=1 𝑎𝑖 − 𝑏12 +𝑏𝑖 𝑥 +𝑥 4 [−5, 5] 0.00030
𝑖 𝑖 3 4

𝐹 − 16(𝑥) = 4𝑥21 − 2.1𝑥41 + 31 𝑥61 + 𝑥1 𝑥2 − 4𝑥22 + 4𝑥42 2 [−5, 5] −1.0316

( )2 ( )
5.1 2
𝐹 − 17(𝑥) = 𝑥2 − 4𝜋 2
𝑥1 + 𝜋5 𝑥1 − 6 + 10 1 − 8𝜋 1
cos 𝑥1 + 10 2 [−5, 5] 0.398
𝐹 − 18(𝑥) = 𝐴(𝑥) × 𝐵(𝑥)
( )2 ( )
𝐴(𝑥) = 1 + 𝑥1 + 𝑥2 + 1 19 − 14𝑥1 + 3𝑥21 − 14𝑥2 + 6𝑥1 𝑥2 + 3𝑥22 2 [−2, 2] 3
( )2 ( )
𝐵(𝑥) = 30 + 2𝑥1 − 3𝑥2 (× 18 − 32𝑥1 + 12𝑥21 +)48𝑥2 − 36𝑥1 𝑥2 + 27𝑥22
∑4 ∑3 ( )2
𝐹 − 19(𝑥) = − 𝑖=1 𝑐𝑖 exp − 𝑗=1 𝑎𝑖𝑗 𝑥𝑗 − 𝑝𝑖𝑗 3 [1, 3] −3.86
∑4 ( ∑ ( )2 )
6
𝐹 − 20(𝑥) = − 𝑖=1 𝑐𝑖 exp − 𝑗=1 𝑎𝑖𝑗 𝑥𝑗 − 𝑝𝑖𝑗 6 [0, 1] −3.32
∑5 [ ( )( )𝑇 ]−1
𝐹 − 21(𝑥) = − 𝑖=1 𝑋 − 𝑎𝑖 𝑋 − 𝑎𝑖 + 𝑐𝑖 4 [0, 10] −10.1532
∑7 [ ( )( )𝑇 ]−1
𝐹 − 22(𝑥) = − 𝑖=1 𝑋 − 𝑎𝑖 𝑋 − 𝑎𝑖 + 𝑐𝑖 4 [0, 10] −10.4028
[
∑10 ( )( )𝑇 ]−1
𝐹 − 23(𝑥) = − 𝑖=1 𝑋 − 𝑎𝑖 𝑋 − 𝑎𝑖 + 𝑐𝑖 4 [0, 10] −10.5363

∙ Step 1: Initialize the positions of the particles by Logistic chaos ∙ Step 9: Determine whether the number of iterations reaches the
mapping and assign values to each parameter. maximum. If yes, outputting the result; otherwise return to step
∙ Step 2: Calculate the fitness value of the particle and save the 3.
local optimal Pbest (i) and the global optimal position gbest.
4.4.6. Algorithm performance testing
∙ Step 3: Calculate 𝑓𝑤𝑜𝑟𝑠𝑡 , 𝑓𝑏𝑒𝑠𝑡 and 𝑓𝑎𝑣𝑔 of the particles and update
In this paper, the CEC-2005 benchmark function sets are selected
the adaptive control coefficients w, 𝑐1 , 𝑐2 .
to test the optimization performance of IGSAPSO. Table 1 gives the
∙ Step 4: Calculate the mass 𝑀𝑖 (k), gravitational constant G(k),
definition formula, dimensions, range of values and the theoretical
combined force 𝐹𝑖 (k) and acceleration 𝑎𝑖 (k) of each particle. optimal value of 23 functions. In Table 1, F-1∼F-7 are Unimodal
∙ Step 5: Update the particle velocity and check if the velocity benchmark functions, F-8∼F-13 are Multimodal benchmark functions
exceeds the constraint limit, and correct for particles that exceed and F-14∼F-23 are fixed-dimension multimodal benchmark function.
the velocity limit. In order to verify the feasibility and optimization performance of
∙ Step 6: Update the particle positions and correct the positions of the algorithm proposed in this paper. It is compared with PSO, GSA,
GSAPSO [25], HGSPSO [28], SBG-GSAPSO [27], MGSAPSO [31], IS-
the particles that exceed the constraint limits.
GSAPSO [36] algorithm for the benchmark test function search and
∙ Step 7: Update the fitness values of the particles and save the local
comparison experiments. The experimental parameters of all algo-
optimal and global optimal positions of the particle population.
rithms are particle population size n = 100, the maximum number of
∙ Step 8: The elite strategy is used to generate new merit-seeking iterations K = 1000, and the dimensionality d = 50 of the multidi-
populations, calculate the fitness values of the new populations mensional test function. The specific parameters of each algorithm are
and save the locally optimal and globally optimal positions. shown in Table 2.

8
K. Pan et al. International Journal of Electrical Power and Energy Systems 157 (2024) 109766

Table 2
Parameters of each algorithm.
Algorithms Parameter settings
PSO 𝑐1 = 𝑐2 = 1.5, 𝑤 = 0.8
GSA 𝐺0 = 100, 𝜓 = 0.2
GSAPSO [25] 𝑐1 = 𝑐2 = 2, 𝑤 = 0.5, 𝐺0 = 100, 𝜓 = 20
HGSPSO [28] 𝑐3 = 𝑐6 = 2.5, 𝑐4 = 𝑐5 = 0.5, 𝑤 = 0.9, 𝐺0 = 1, 𝜓 = 20
SBG-GSAPSO [27] 𝑐1 = 1.2, 𝑐2 = 1.5, 𝐺0 = 100, 𝜓 = 20
1 2 1 2
MGSAPSO [31] 𝑐max = 𝑐max = 2.5, 𝑐min = 𝑐min = 0.5, 𝑤max = 0.9, 𝑤min = 0.4, 𝐺0 = 100, 𝜓 = 0.2
IS-GSAPSO [36] 𝑐1 = 𝑐2 = 2, 𝑤≇ = 0.9, 𝛼𝑤 = 0.1, 𝐺0 = 90, 𝜓 = 0.1
1 2 1 2
IGSAPSO 𝑐max = 𝑐max = 2.5, 𝑐min = 𝑐min = 0.6, 𝑤max = 0.9, 𝑤min = 0.4, 𝐺0 = 100, 𝜓 = 20

Table 3
Comparison of IGSAPSO with other algorithms on CEC-2005(Unimodal benchmark function).
Function PSO GSA GSA-SPO HGSPSO SBG-GSAPSO MGSAPSO IS-GSAPSO IGSAPSO
Min 1.43E+00 2.55E−01 8.83E−03 3.64E−06 8.92E−07 9.50E−11 1.98E−02 0.00E+00
Mean 3.11E+00 3.12E−01 1.27E−02 1.63E−03 2.13E−01 8.82E−05 3.70E−02 0.00E+00
F-1
SD 1.10E+00 2.09E−02 2.28E−03 3.71E−03 3.26E−01 1.33E−04 1.09E−02 0.00E+00
Rank 8 7 4 3 6 2 5 1
Min 1.07E+01 2.89E+00 4.24E+00 0.00E+00 5.31E−01 1.40E−01 6.45E−01 0.00E+00
Mean 1.96E+01 3.09E+00 6.68E+00 2.74E−04 1.95E+00 2.25E−01 1.04E+00 0.00E+00
F-2
SD 4.97E+00 1.27E−01 1.76E+00 1.48E−03 8.23E−01 3.34E−02 2.03E−01 0.00E+00
Rank 8 6 7 2 5 3 4 1
Min 6.06E+00 2.82E+01 1.64E+00 4.04E−05 2.19E+02 3.10E−08 7.26E−02 0.00E+00
Mean 1.14E+02 4.22E+01 3.13E+00 1.37E−02 7.76E+02 1.25E−03 1.69E−01 0.00E+00
F-3
SD 3.52E+00 8.89E+00 6.71E−01 2.13E−02 4.22E+02 1.73E−03 6.33E−02 0.00E+00
Rank 6 7 5 3 8 2 4 1
Min 1.30E+01 1.53E−01 4.74E+00 8.04E−05 4.47E+00 8.98E−04 3.20E−02 0.00E+00
Mean 1.63E+01 1.73E−01 9.87E+00 1.67E−03 7.74E+00 5.02E−03 5.37E−02 0.00E+00
F-4
SD 2.06E+00 7.13E−03 2.85E+00 6.70E−03 1.77E+00 2.57E−03 1.08E−02 0.00E+00
Rank 8 4 7 2 6 3 5 1
Min 1.48E+02 7.24E+01 4.45E+01 4.87E+01 4.33E+01 4.87E+01 5.07E+01 0.00E+00
Mean 3.13E+01 7.96E+01 5.00E+01 4.90E+01 9.30E+01 4.88E+01 5.21E+01 0.00E+00
F-5
SD 1.30E+02 2.67E+00 6.17E+00 4.45E−01 5.85E+01 6.16E−02 8.46E−01 0.00E+00
Rank 8 6 4 3 7 2 5 1
Min 8.73E+01 2.45E−01 1.76E−01 6.46E+00 8.38E−17 2.57E−07 3.93E+00 1.76E−14
Mean 3.01E+02 3.24E−01 3.42E−01 9.30E+00 3.20E−01 1.75E−05 5.32E+00 8.83E−06
F-6
SD 1.20E+02 2.91E−02 7.31E−02 1.55E+00 7.49E−01 3.68E−05 8.34E−01 4.75E−05
Rank 8 4 5 7 3 2 6 1
Min 2.56E−02 9.54E−02 2.24E−01 7.39E−05 1.40E−02 7.98E−03 1.19E−03 2.66E−05
Mean 6.06E−02 2.25E−01 5.08E−01 5.00E−03 2.73E−02 1.87E−02 2.00E−02 7.39E−04
F-7
SD 2.12E−02 3.63E−02 1.21E−01 3.52E−03 8.36E−03 7.42E−03 1.01E−02 6.06E−04
Rank 6 7 8 2 5 3 4 1
Overall ranking 7 6 5 3 5 2 4 1
(Average ranking number) (7.4) (5.9) (5.7) (3.1) (5.7) (2.4) (4.7) (1)

In order to compare the optimization performance of various algo- F-14 through F-23 are fixed-dimensional multipeak functions, and
rithms fairly and impartially, all algorithms were simulated 30 times the minimum of the IGSAPSO algorithm is better than or equal to the
under Intel(R) Core(TM) i5-8400 CPU @ 2.80 GHz, 8.00G RAM, win- other algorithms, but the average of F-14, F-15, and F-20 through F-23
dows 10 system and Matlab R2020a, and then the average value(Mean) is optimal. The standard deviation is also optimal except for F-18.
and standard deviation(SD) were calculated. The test results for the Fig. 10 plots the convergence curves of some of the test func-
function F-1∼F-23 are shown in Table 3 ∼ Table 5. tions, from which it can be seen that IGSAPSO is able to converge
Table 3 shows the optimality finding results for the unimodal func- to the near-optimal solution in less than 40 iterations on the single-
tions F-1∼F-7. The unimodal function has only one global optimal peak functions F-1 to F-4 and the multi-peak functions F-9 to F-11,
solution, which is suitable for testing the exploitation capability of which is significantly better than the other algorithms in terms of the
algorithm’s convergence speed. IGSAPSO also converges to the optimal
the algorithm. It is evident from Table 3 that the minimum, mean
value after 400 iterations on the fixed-dimensional multi-peak functions
and standard deviation values all are optimum for proposed IGSAPSO
F-21 to F-23. These results demonstrate that IGSAPSO has excellent
than the other algorithms. This indicates that IGSAPSO has excellent
local development capability, global exploration capability and very
exploited capability. For functions F-1∼F-5, the IGSAPSO proposed in
fast convergence speed.
this paper reaches the global optimization and has no deviation in
Box plots can show the distribution and degree of fluctuation of
multiple tests, which indicates that the proposed algorithm has strong target values obtained from multiple tests, and they can provide a
stability and robustness. wealth of information about the test sample (like the minimum, me-
Multimodal functions with many local optimas are suitable for dian, median upper or lower quartile, maximum). Fig. 11 shows the
testing the global exploration capability of an algorithm. The results box plots of eight algorithms for 23 classical benchmark functions, and
of the multimodal function tests F-8 to F-13 are shown in Table 4 it can be found that the distribution of the objective values of IGSAPSO
IGSAPSO converges to the optimal or near-optimal values in the F-8 is significantly smaller than that of other algorithms, which proves that
to F-11 function tests, and also performs well in the F-12 and F-13 it has strong robustness.
function solving. It shows that the improved algorithm has strong Combining classical benchmark function test results, convergence
global exploration capability without falling into local minima. characterization, and box plot analysis, IGSAPSO performs well on

9
K. Pan et al. International Journal of Electrical Power and Energy Systems 157 (2024) 109766

Table 4
Comparison of IGSAPSO with other algorithms on CEC-2005(Multimodal benchmark function).
Function PSO GSA GSA-SPO HGSPSO SBG-GSAPSO MGSAPSO IS-GSAPSO IGSAPSO
Min −1.29E+04 −5.78E+03 −1.32E+04 −9.14E+03 −1.31E+04 −2.06E+04 −1.13E+04 −2.09E+04
Mean −1.08E+04 −3.84E+03 −1.02E+04 −4.65E+03 −1.08E+04 −1.82E+04 −8.76E+03 −2.09E+04
F-8
SD 1.24E+03 6.01E+02 1.15E+03 1.44E+03 1.11E+03 1.39E+03 1.26E+03 2.92E−03
Rank 3 8 5 7 4 2 6 1
Min 7.40E+01 5.21E+01 6.84E+01 1.29E−05 2.89E+01 0.00E+00 2.29E+00 0.00E+00
Mean 1.14E+02 6.19E+01 1.13E+02 2.13E−04 5.08E+01 3.05E−01 8.65E+00 0.00E+00
F-9
SD 3.08E+01 4.22E+00 2.35E+01 2.65E−04 1.48E+01 8.91E−02 1.55E+01 0.00E+00
Rank 8 6 7 2 5 3 4 1
Min 6.38E+00 5.62E−01 1.74E−01 8.79E−05 4.91E+00 7.09E−03 4.53E−02 8.88E−16
Mean 7.70E+00 1.55E+00 2.03E+00 3.35E−01 6.50E+00 1.15E−02 7.00E−02 8.88E−16
F-10
SD 8.40E−01 1.15E+00 8.19E−01 9.07E−01 8.43E−01 2.09E−03 1.58E−02 9.86E−32
Rank 8 5 6 4 7 2 3 1
Min 2.18E+00 1.24E−02 2.82E−02 4.02E−08 1.25E+00 0.00E+00 7.39E−04 0.00E+00
Mean 3.88E+00 6.91E−02 4.51E−02 4.18E−02 2.26E+00 3.00E−08 1.55E−03 0.00E+00
F-11
SD 9.63E−01 1.77E−01 1.13E−02 1.26E−01 9.29E−01 1.60E−07 5.17E−04 0.00E+00
Rank 8 6 5 4 7 2 3 1
Min 4.65E+00 1.71E−03 7.52E+00 1.07E−01 2.13E+00 1.04E−08 1.22E−01 3.27E−16
Mean 1.07E+01 3.40E−02 1.25E+01 2.91E−01 4.98E+00 5.07E−08 2.12E−01 9.58E−05
F-12
SD 3.97E+00 5.57E−02 3.62E+00 1.24E−01 1.94E+00 2.45E−08 4.43E−02 2.72E−04
Rank 7 3 8 5 6 1 4 2
Min 5.69E+01 3.67E−02 4.66E+01 4.32E+00 3.81E+01 1.74E−07 1.91E+00 8.55E−15
Mean 9.44E+01 9.68E−02 7.98E+01 4.95E+00 5.96E+01 1.20E−06 3.05E+00 1.05E−02
F-13
SD 2.88E+01 2.46E−01 1.55E+01 5.74E−01 1.21E+01 7.03E−07 6.14E−01 4.00E−02
Rank 8 3 7 5 6 1 4 2
Overall ranking 8 5 7 4 6 2 3 1
(Average ranking number) (7.0) (5.2) (6.3) (4.5) (5.8) (1.8) (4.0) (1.3)

Table 5
Comparison of IGSAPSO with other algorithms on CEC-2005(Fixed-dimension multimodal benchmark function).
Min 9.98E−01 9.98E−01 9.98E−01 9.98E−01 9.98E−01 9.98E−01 9.98E−01 9.98E−01
Mean 1.72E+00 7.42E+00 9.20E+00 1.39E+00 9.01E+00 9.98E−01 9.98E−01 9.98E−01
F-14
SD 1.32E+00 4.39E+00 6.53E+00 9.43E−01 5.19E+00 3.66E−10 1.21E−04 1.22E−16
Rank 5 6 7 4 8 2 3 1
Min 3.07E−04 1.85E−03 3.33E−04 3.11E−04 3.07E−04 3.29E−04 4.31E−04 3.07E−04
Mean 2.76E−03 7.36E−03 4.01E−03 5.80E−04 4.71E−04 4.37E−04 8.01E−04 3.73E−04
F-15
SD 6.14E−03 2.68E−03 6.68E−03 5.01E−04 3.50E−04 8.58E−05 1.77E−04 7.44E−05
Rank 6 8 7 4 3 2 5 1
Min −1.03E+00 −1.03E+00 −1.03E+00 −1.03E+00 −1.03E+00 −1.03E+00 −1.03E+00 −1.03E+00
Mean −1.03E+00 −1.03E+00 −1.03E+00 −1.02E+00 −1.03E+00 −1.03E+00 −1.03E+00 −1.03E+00
F-16
SD 6.35E−16 2.66E−04 2.67E−05 9.55E−03 4.62E−16 2.01E−04 4.17E−04 4.33E−16
Rank 3 5 7 8 2 6 4 1
Min 3.98E−01 3.98E−01 3.98E−01 3.98E−01 3.98E−01 3.98E−01 3.98E−01 3.98E−01
Mean 3.98E−01 3.98E−01 3.98E−01 3.98E−01 3.98E−01 3.98E−01 3.98E−01 3.98E−01
F-17
SD 0.00E+00 1.25E−04 1.63E−05 7.50E−05 0.00E+00 1.27E−04 1.65E−04 0.00E+00
Rank 1 4 2 3 1 5 6 1
Min 3.00E+00 3.00E+00 3.00E+00 3.00E+00 3.00E+00 3.00E+00 3.00E+00 3.00E+00
Mean 3.00E+00 3.03E+00 3.00E+00 3.00E+00 3.00E+00 3.01E+00 3.01E+00 3.00E+00
F-18
SD 9.93E−16 3.85E−02 1.26E−03 3.58E−04 2.87E−15 6.52E−03 5.89E−03 5.96E−15
Rank 1 8 4 5 2 6 7 3
Min −3.86E+00 −3.85E+00 −3.86E+00 −3.69E+00 −3.86E+0 −3.86E+00 −3.86E+00 −3.86E+00
Mean −3.86E+00 −3.42E+00 −3.84E+00 −3.10E+00 −3.86E+00 −3.86E+00 −3.86E+00 −3.86E+00
F-19
SD 1.97E−03 4.58E−01 8.63E−02 4.07E−01 2.23E−15 1.91E−03 2.52E−03 1.86E−15
Rank 4 7 6 8 2 3 5 1
Min −3.32E+00 −3.12E+00 −3.32E+00 −2.77E+00 −3.32E+00 −3.31E+00 −3.30E+00 −3.32E+00
Mean −3.25E+00 −2.78E+00 −3.07E+00 −1.64E+00 −3.30E+00 −3.28E+00 −3.20E+00 −3.31E+00
F-20
SD 7.71E−02 2.61E−01 2.02E−01 5.42E−01 4.43E−02 4.57E−02 5.54E−02 3.56E−02
Rank 4 7 6 8 2 3 5 1
Min −1.02E+01 −1.00E+01 −1.01E+01 −1.02E+01 −1.02E+01 −1.01E+01 −1.01E+01 −1.02E+01
Mean −6.07E+00 −8.29E+00 −6.67E+00 −8.09E+00 −8.31E+00 −8.05E+00 −8.58E+00 −1.02E+01
F-21
SD 3.62E+00 2.15E+00 3.30E+00 3.58E+00 2.90E+00 2.47E+00 2.11E+00 1.11E−07
Rank 8 3 7 2 4 6 2 1
Min −1.04E+01 −9.99E+00 −1.04E+01 −1.04E+01 −1.04E+01 −1.04E+01 −1.04E+01 −1.04E+01
Mean −8.99E+00 −8.90E+00 −7.33E+00 −8.51E+00 −9.46E+00 −9.11E+00 −9.50E+00 −1.04E+01
F-22
SD 2.86E+00 1.71E+00 3.80E+00 3.55E+00 2.42E+00 2.24E+00 1.60E+00 1.86E−05
Rank 5 6 8 7 3 4 2 1
Min −1.05E+01 −1.05E+01 −1.05E+01 −1.05E+01 −1.05E+01 −1.05E+01 −1.04E+01 −1.05E+01
Mean −7.58E+00 −6.82E+00 −8.74E+00 −7.77E+00 −1.03E+01 −9.57E+00 −1.02E+01 −1.05E+01
F-23
SD 3.66E+00 3.98E+00 3.01E+00 3.51E+00 1.46E+00 2.00E+00 2.30E−01 2.53E−15
Rank 7 8 5 6 2 4 3 1
Overall ranking 5 8 7 6 2 3 4 1
(Average ranking number) (4.4) (6.2) (5.9) (5.8) (2.9) (4.1) (4.2) (1.2)

10
K. Pan et al. International Journal of Electrical Power and Energy Systems 157 (2024) 109766

Fig. 10. Time-of-use tariff of general industrial in Beijing.

almost all benchmark function tests. This demonstrates the good perfor- the base load of 96 time periods is obtained by fitting. It is assumed that
mance and feasibility of IGSAPSO. After conducting tests and observing 500 EVs of the same type are available for dispatch in the region, and
the positive impact of IGSAPSO on the classical benchmark function, the specific parameters of the EVs are shown in Table 7 . The maximum
the applicability of the algorithm to real engineering problems will be load capacity of the regional feeder is 6400 kVA. The EV starting load
examined in the next section. 𝑎𝑐 (0.5, 0.12 ). The load state that
state satisfies a normal distribution 𝑁𝑠𝑜𝑐
the user expects to leave the grid is set to no less than 0.9; the initial
5. Case study weights 𝛼 = 𝛽 = 0.5, and the parameters of the algorithm are set as in
Section 4.4.6. The division of the time-of-use tariff is based on the peak
5.1. Parameter settings and valley tariffs for general industry in Beijing, as shown in Fig. 12
(see Table 6).
Taking a residential area as an example, its feeder base load refers Based on the above theory and data, the daily charging load
to the base load distribution in literature [24], as shown in Table 6, and distribution of EVs in the region can be obtained by simulating the

11
K. Pan et al. International Journal of Electrical Power and Energy Systems 157 (2024) 109766

Fig. 10. (continued).

Table 6 EVs charge/discharge optimization model using IGSAPSO are ¥2184.3,

Foundation load.
642.7 kW and 3.01 × 106 kW2 respectively for the same overall charging
Time Load/kW Time Load/kW Time Load/kW
power of the EVs. All three metrics are the smallest of all algorithm
01:00 1625.3 09:00 3018.4 17:00 2321.9 solutions, and the charging costs are 18.7%, 2.9% and 12.5% lower
02:00 1741.4 10:00 3250.6 18:00 3018.4
than PSO, GSA and GSAPSO, respectively; the load peak-to-valley
03:00 1973.6 11:00 3482.8 19:00 3482.8
04:00 2205.8 12:00 3250.6 20:00 3715.0 differences are 9.7%, 13.4% and 9.4% lower; the load variance is 5.9%,
05:00 2321.9 13:00 3018.4 21:00 3018.4 27.6% and 5.9% lower. The solution results of the IGSAPSO algorithm
06:00 2554.1 14:00 2879.1 22:00 2554.1 are also significantly better compared to other improved hybrid particle
07:00 2670.2 15:00 2876.3 23:00 2089.7
swarm optimization and gravitational search algorithms. In particular,
08:00 2786.3 16:00 2438.0 24:00 1857.5
although the total cost of EV charging when solved using the IS-
GSAPSO algorithm is slightly lower than that of IGSAPSO (0.1%),
Table 7 the distribution network load variance and peak-to-valley difference
Foundation load.
are substantially higher than that of IGSAPSO (5.9%). In summary,
Parameters Value
the use of IGSAPSO algorithm for solving such optimization problems
Battery capacity 35 kWh can reduce the charging cost to a greater extent and have a stronger
SOC range [0.3,1]
suppression effect on the load fluctuation of the grid.
Charging/Discharging power 7 kW
Charging/Discharging efficiency 0.9 Fig. 13 illustrates the optimization-seeking iteration curves of the
Battery replacement cost 35,000¥ eight algorithms when scheduling the same EV. It can be seen that
IGSAPSO has the lowest objective function value at the convergence
stage. Although IGSAPSO converges at the 800th iteration, the fitness
value already exceeds the other algorithms after 180 iterations. The
charging and discharging behavior of 500 EVs in one day using Monte
proposed variant of the particle swarm and gravitational search algo-
Carlo (MC).
rithm shows superior characteristics through box-and-line plot analysis.
It can be seen that IGSAPSO effectively handles the optimization pro-
5.2. Analysis of scheduling results of different algorithms
cess and the positive impact of the modification of the hybrid algorithm
The results of solving the EV charging and discharging optimization is quite evident.
model by eight algorithms are given in Table 8. The starting 𝑆𝑂𝐶𝑎𝑐
mean value, off-grid 𝑆𝑂𝐶𝑑𝑒 mean value, total charging cost, and the 5.3. Scheduling results for different charging strategies
peak-to-valley difference and variance of the grid load curve for 500
EVs under different algorithm scheduling are calculated, respectively, To analyze the optimization performance of solving the EV ordered
while ensuring the consistency of each parameter. charging and discharging scheduling problem using IGSAPSO. The
As can be seen from Table 8, the total cost of charging, the peak- total cost of EV charging, the variance of grid load, and the peak-
to-valley difference in grid load and the variance when solving the to-valley difference are used as evaluation indicators. The scheduling

12
K. Pan et al. International Journal of Electrical Power and Energy Systems 157 (2024) 109766

Fig. 11. Boxplot analysis on CEC-2005.

results of disordered charging, ordered charging and ordered charg- are reduced by 31.9%, 57.5% and 75.3%, respectively. In addition, the
ing/discharging are compared, as shown in the Table 9. ordered charging and discharging of EVs connected to the regional grid
From Table 9, it can be seen that the total cost of charging under based on the initial regional load reduce the peak-to-valley difference
the ordered charging and discharging schedule of EVs is reduced by and the fluctuation variance of the regional grid load by 69% and
68.1% compared to the disordered charging, while the peak–valley 90.3%, respectively.
difference and variance of the grid load are reduced by 80.8% and Fig. 14 shows the daily load curves under different scheduling
95.8%, respectively. Compared with ordered charging, the total cost strategies. It can be seen that large-scale electric vehicles connected to
of charging, the peak–valley difference and the variance of grid load the grid for disordered charging will cause the phenomenon of ‘‘peak

13
K. Pan et al. International Journal of Electrical Power and Energy Systems 157 (2024) 109766

Fig. 11. (continued).

5.4. Scheduling results with different optimization weights

The dispatch center can set the optimization weights according to

the operation status of the grid and the charging demand of the users
to balance the rights and interests of both parties. Taking 500 EVs
in the region as an example, we analyze the optimization results of
charging and discharging when EV users participate in dispatching with
optimization weights of 𝛼∕𝛽 = 0.2/0.8, 𝛼∕𝛽 = 0.4/0.6, 𝛼∕𝛽 = 0.5/0.5,
𝛼∕𝛽 = 0.6/0.4 and 𝛼∕𝛽 = 0.8/0.2, respectively. Fig. 15 illustrates the
distribution of daily load curves of the grid with different optimization
weights, and the specific optimization results are shown in Table 10.
Table 10 shows that the charging cost is significantly lower when
EV users participate in the ordered charging and discharging schedule
compared to the unordered charging. The charging cost decreases con-
Fig. 12. Time-of-use tariff of general industrial in Beijing. tinuously with the increase of the optimization weight 𝛽. The charging
cost decreases significantly when the weight 𝛽 increases from 0.6 to
0.8, indicating that the user benefits from participating in the ordered
charging and discharging schedule with the optimization weight. In
addition, as the load variance optimization weight 𝛼 increases, the
on peak’’ and exceed the regional load limit at the peak load value. This
phenomenon can be avoided by arranging EV charging in an ordered scheduling system arranges EVs to charge during off-peak periods and
manner, but a new load peak will be formed between 3:00 and 7;00 discharge during off-peak periods in order to smooth out the load
am. Moreover, due to the incentive of time-of-use tariff, users choose fluctuation of the grid, which causes the charging cost to increase
to charge at low tariff time, which will cause drastic fluctuation of grid subsequently. Moreover, the charging power also increases with the
load at 7:00 am and 11:00 pm, thus affecting the power quality of the increase of 𝛼, which is due to the fact that the smoothing effect of
grid. The overall load curve under ordered charging and discharging EV charging more power on the grid load is better than the additional
schedule is relatively flat, with a small trough area formed between charging cost.
6:00 and 9:00 pm. This is mainly formed by the large number of EV Fig. 16 illustrates the daily load curve of the grid under different
users driven by peak tariffs to discharge during this time to obtain high weight scheduling. It can be seen that when the optimized weight
revenue. Overall, using IGSAPSO to solve the EV ordered charging and 𝛽 = 0.8 of charging cost, although the charging cost of users is the
discharging scheduling problem can smooth out the load fluctuation of lowest, there will be over-cutting and over-filling of peaks and valleys,
the grid and play the role of peak shaving and valley filling for the grid. resulting in new peaks and valleys in the grid. Furthermore, as the

14
K. Pan et al. International Journal of Electrical Power and Energy Systems 157 (2024) 109766

Fig. 13. Convergence curve and box plot analysis of electric vehicle charge–discharge optimization problems.

Table 8
Scheduling results of each algorithm.
Indicators Average value(𝑆𝑂𝐶𝑎𝑐 ) Average value(𝑆𝑂𝐶𝑑𝑒 ) Total cost of charging/¥ Load peak–valley difference/kW Load variance/kW2
PSO 2686.72 711.63 3.20 × 106
GSA 2249.37 742.21 4.16 × 106
GSAPSO [25] 2497.64 709.65 3.20 × 106
HGSPSO [28] 2421.57 658.50 3.05 × 106
0.5 0.91
SBG-GSAPSO [27] 2262.44 674.47 3.13 × 106
MGSAPSO [31] 2206.46 663.38 3.08 × 106
IS-GSAPSO [36] 2181.96 710.23 3.19 × 106
IGSAPSO 2184.35 642.69 3.01 × 106

Table 9
Scheduling results for different charging strategies.
Strategy Total cost of charging/¥ Load peak–valley difference/kW Load variance/kW2
Foundation load – 2074.21 3.10 × 107
Disordered charging 6848.83 3357.58 7.21 × 107
Ordered charging 3208.84 1513.31 1.22 × 107
Ordered charging and discharging 2184.35 642.69 3.01 × 106

Fig. 14. Load curves for different charging strategies.

Table 10
Scheduling results for different charging strategies.
Charging mode 𝛼∕𝛽 Average value Average value Total cost of Cost of uncontrolled charging Load peak–valley Load variance/
(SOCac) (SOCde) charging/¥ for the same charge/¥ difference/kW kW2
Disordered charging – 0.91 6848.83 – 3357.58 7.21 × 107
0.2/0.8 0.90 60.45 6471.11 1537.67 2.22 × 107
0.4/0.6 0.5 0.91 1632.27 6632.89 839.34 5.68 × 106
Ordered charging and discharging 0.5/0.5 0.91 2184.35 6632.89 642.69 3.01 × 106
0.6/0.4 0.92 2568.65 6794.67 560.22 1.98 × 106
0.8/0.2 0.93 3189.43 6956.44 486.36 1.07 × 106

15
K. Pan et al. International Journal of Electrical Power and Energy Systems 157 (2024) 109766

Fig. 15. Optimization results with different weights.

Fig. 16. Daily load profile of the grid at different participation levels.

Table 11 degree. In addition, Fig. 16 also shows that the suppression effect on
Scheduling results for different charging strategies.
the grid load curve increases with the increase of user participation,
Participation Total cost of Load peak–valley Load variance/kW2 and the effect of peak and valley reduction is more obvious. The more
charging/¥ difference/kW
active EV users respond to V2G mode and the more EVs participate in
20% 4936.07 1191.2 1.21 × 107
dispatching, the stronger the suppression effect on grid load will be,
40% 3671.72 979.15 5.11 × 106
60% 2708.16 925.96 4.24 × 106 and the economic benefits of users will be gradually increased.
80% 2460.52 803.10 3.58 × 106
100% 2184.35 642.69 3.01 × 106 6. Conclusions

In this paper, we start with the stable operation of the grid and
optimization weight of cost decreases, the optimization weight of load the practical benefits of electric vehicle users. A multi-objective opti-
variance increases and the suppression of load fluctuation of the grid mal scheduling model is proposed to achieve the minimum grid load
is significantly enhanced. variance and the minimum charging cost (including battery recession
cost). To solve this type of optimization problem, a hybrid algorithm
5.5. Optimization results for different number of EVs involved in scheduling combining an improved PSO and GSA is proposed. The algorithm in-
troduces the core mechanism of GSA in which gravitational interaction
To analyze the impact of user participation in ordered charging and exists between celestial bodies into the PSO. Moreover, the algorithm
discharging scheduling positivity on the grid load suppression effect. is improved by using chaotic initialization population, adaptive inertia
The participation degree is defined as the ratio of EVs responding to weight coefficients, learning factors and elite learning strategy, which
V2G mode to the number of all EVs in the region that need to be effectively improves the algorithm’s performance of finding the best.
charged. A total of five cases with participation degrees of 20%, 40%, The performance of the algorithm is tested by CEC2005 benchmarking
60%, 80% and 100% are set for simulation. The optimization weights of function sets, and the proposed IGSAPSO algorithm outperforms PSO,
the participating dispatched vehicles are all set to 𝛼 = 𝛽 = 0.5. Table 11 GSA and their hybrid algorithms. The proposed optimization scheme
and Fig. 16 give the optimization results and the daily grid load curve is tested for scheduling under an ordered charging and discharging
distribution for different participation degrees. As shown in Table 11, strategy. The simulation results show that the total charging cost
the total charging cost, the peak-to-valley difference and variance of the and grid load variance are reduced by 18.7% and 5.9%, respectively,
grid load in the region decrease with the increase of the participation when the optimization model is solved using IGSAPSO compared to

16
K. Pan et al. International Journal of Electrical Power and Energy Systems 157 (2024) 109766

the standard PSO. It shows that this ordered charging and discharg- [9] Chen X, Leung K-C. Non-cooperative and cooperative optimization of
ing optimization scheduling scheme can effectively reduce the total scheduling with vehicle-to-grid regulation services. IEEE Trans Veh Technol
2019;69(1):114–30. http://dx.doi.org/10.1109/TVT.2019.2952712.
charging cost of EV users and improve the operational stability of the
[10] Bouhouras AS, Kothona D, Gkaidatzis PA, Christoforidis GC. Distribution net-
grid. Furthermore, when the optimization weights are all set to 0.5, work energy loss reduction under EV charging schedule. Int J Energy Res
the ordered charging and discharging scheduling strategy can reduce 2022;46(6):8256–70. http://dx.doi.org/10.1002/er.7727.
the total charging cost of users by 68.1% and the grid load variance [11] He Y, Venkatesh B, Guan L. Optimal scheduling for charging and discharging of
by 95.8% compared with the unordered charging strategy. Compared electric vehicles. IEEE Trans Smart Grid 2012;3(3):1095–105. http://dx.doi.org/
10.1109/TSG.2011.2173507.
with ordered charging, the reduction of charging cost and load vari-
[12] Cao Y, Tang S, Li C, Zhang P, Tan Y, Zhang Z, Li J. An optimized EV
ance reaches 31.9% and 75.3%, respectively, reflecting the excellent charging model considering TOU price and SOC curve. IEEE Trans Smart Grid
performance of the scheduling scheme. The multi-objective ordered 2011;3(1):388–93. http://dx.doi.org/10.1109/TSG.2011.2159630.
charging and discharging scheme described in this paper provides a [13] Zhou K, Cheng L, Lu X, Wen L. Scheduling model of electric vehicles charg-
new way to solve the charging management of EVs in the region’s ing considering inconvenience and dynamic electricity prices. Appl Energy
residential communities and maintain the power grid’s stable operation. 2020;276:115455. http://dx.doi.org/10.1016/j.apenergy.2020.115455.
[14] Wei D, Zhang C, Sun B, Cui N. A time-of-use price based multi-objective optimal
It provides new ideas for effective management control of large-scale
dispatching for charging and discharging of electric vehicles. Power Syst Technol
electric vehicle charging and discharging in an ordered manner in the 2014;38(11):2972–7. http://dx.doi.org/10.13335/j.1000-3673.pst.2014.11.005.
future. [15] Chen Q, Wang W, Wang H, Dong Y, He S. Information gap-based coordination
scheme for active distribution network considering charging/discharging opti-
Formatting of funding sources mization for electric vehicles and demand response. Int J Electr Power Energy
Syst 2023;145. http://dx.doi.org/10.1016/j.ijepes.2022.108652.
[16] Ma X, Wang C, Hong X, Wang H, Li Y. Optimal scheduling of charging
This research did not receive any specific grant from funding agen- and discharging of electric vehicle based on real time price and economic
cies in the public, commercial, or not-for-profit sectors. dispatch model. Trans China Electrotech Soc 2016;31:190–202. http://dx.doi.
org/10.19595/j.cnki.1000-6753.tces.2016.s1.025.
CRediT authorship contribution statement [17] Srilakshmi E, Singh SP. Energy regulation of EV using MILP for optimal operation
of incentive based prosumer microgrid with uncertainty modelling. Int J Electr
Power Energy Syst 2022;134. http://dx.doi.org/10.1016/j.ijepes.2021.107353.
Kui Pan: Conceptualization, Methodology, Software, Formal analy- [18] Merhy G, Nait-Sidi-Moh A, Moubayed N. Control, regulation and optimization
sis, Writing – original draft. Chuan-Dong Liang: Data curation, Formal of bidirectional energy flows for electric vehicles’ charging and discharging.
analysis, Software, Writing – original draft. Min Lu: Conceptualization, Sustainable Cities Soc 2020;57:102129. http://dx.doi.org/10.1016/j.scs.2020.
Resources, Supervision, Writing – review & editing. 102129.
[19] Lin B, Ghaddar B, Nathwani J. Electric vehicle routing with charging/discharging
under time-variant electricity prices. Transp Res C 2021;130:103285. http://dx.
Declaration of competing interest
doi.org/10.1016/j.trc.2021.103285.
[20] Habib HUR, Subramaniam U, Waqar A, Farhan BS, Kotb KM, Wang S. Energy
The authors declare that they have no known competing financial cost optimization of hybrid renewables based V2G microgrid considering multi
interest or personal relationships that could have appeared to influence objective function by using artificial bee colony optimization. IEEE Access
the work reported in this paper. 2020;8:62076–93. http://dx.doi.org/10.1109/ACCESS.2020.2984537.
[21] Poursistani M, Abedi M, Hajilu N, Gharehpetian G. Smart charging of plug-
in electric vehicle using gravitational search algorithm. In: 2014 smart grid
Data availability conference (SGC). IEEE; 2014, p. 1–7. http://dx.doi.org/10.1109/SGC.2014.
7090884.
Data will be made available on request. [22] Vasant P, Marmolejo JA, Litvinchev I, Aguilar RR. Nature-inspired meta-
heuristics approaches for charging plug-in hybrid electric vehicle. Wirel Netw
2020;26:4753–66. http://dx.doi.org/10.1007/s11276-019-01993-w.
References
[23] Zheng Z, Yang S. Particle swarm optimisation for scheduling electric vehicles
with microgrids. In: 2020 IEEE congress on evolutionary computation (CEC).
[1] Anwar MB, Muratori M, Jadun P, Hale E, Bush B, Denholm P, Ma O, Pod-
IEEE; 2020, p. 1–7. http://dx.doi.org/10.1109/CEC48606.2020.9185853.
kaminer K. Assessing the value of electric vehicle managed charging: a review
[24] Liang Z, Chenglong S, Guowei C, Nantian H, Ling L. Two-stage optimization
of methodologies and results. Energy Environ Sci 2022. http://dx.doi.org/10.
strategy for coordinated charging and discharging of EVs based on PSO al-
1039/D1EE02206G.
gorithm. Proc CSEE 2022;(005):042. http://dx.doi.org/10.13334/j.0258-8013.
[2] Das HS, Rahman MM, Li S, Tan C. Electric vehicles standards, charging infras-
pcsee.211150.
tructure, and impact on grid integration: A technological review. Renew Sustain
[25] Mirjalili S, Hashim SZM. A new hybrid PSOGSA algorithm for function optimiza-
Energy Rev 2020;120:109618. http://dx.doi.org/10.1016/j.rser.2019.109618.
tion. In: 2010 international conference on computer and information application.
[3] Jahangir H, Gougheri SS, Vatandoust B, Golkar MA, Ahmadian A, Hajizadeh A.
IEEE; 2010, p. 374–7. http://dx.doi.org/10.1109/ICCIA.2010.6141614.
Plug-in electric vehicle behavior modeling in energy market: A novel deep
learning-based approach with clustering technique. IEEE Trans Smart Grid [26] Jayaprakasam S, Rahim SKA, Leow CY. PSOGSA-Explore: A new hybrid meta-
2020;11(6):4738–48. http://dx.doi.org/10.1109/TSG.2020.2998072. heuristic approach for beampattern optimization in collaborative beamforming.
[4] Zheng Y, Niu S, Shang Y, Shao Z, Jian L. Integrating plug-in electric vehi- Appl Soft Comput 2015;30:229–37. http://dx.doi.org/10.1016/j.asoc.2015.01.
cles into power grids: A comprehensive review on power interaction mode, 024.
scheduling methodology and mathematical foundation. Renew Sustain Energy [27] Hou X, Gao S, Qiu L, Li Z, Zhu R, Lyu S-K. Transmission efficiency optimal design
Rev 2019;112:424–39. http://dx.doi.org/10.1016/j.rser.2019.05.059. of spiral bevel gear based on hybrid PSOGSA (particle swarm optimization-
[5] Hu Z, Song Y, Xu Z, Luo Z, Zhan K, Jia L. Impacts and utilization of electric gravitational search algorithm) method. Appl Sci-Basel 2022;12(19). http://dx.
vehicles integration into power systems. In: Proceedings of the Chinese society of doi.org/10.3390/app121910140.
electrical engineering, Vol. 32. Chinese Society for Electrical Engineering; 2012, [28] Khan TA, Ling SH. A novel hybrid gravitational search particle swarm opti-
p. 1–10. http://dx.doi.org/10.13334/j.0258-8013.pcsee.2012.04.006. mization algorithm. Eng Appl Artif Intell 2021;102. http://dx.doi.org/10.1016/
[6] Tan KM, Ramachandaramurthy VK, Yong JY. Integration of electric vehicles in j.engappai.2021.104263.
smart grid: A review on vehicle to grid technologies and optimization techniques. [29] Li P, Xu W, Zhou Z, Li R. Optimal operation of microgrid based on improved
Renew Sustain Energy Rev 2016;53:720–32. http://dx.doi.org/10.1016/j.rser. gravitational search algorithm. Proc CSEE 2014;34(19):3073–9. http://dx.doi.
2015.09.012. org/10.13334/j.0258-8013.pcsee.2014.19.005.
[7] Wu W, Lin Y, Liu R, Li Y, Zhang Y, Ma C. Online EV charge scheduling based [30] Purcaru C, Precup R-E, Iercan D, Fedorovici L-O, David R-C. Hybrid PSO-GSA
on time-of-use pricing and peak load minimization: Properties and efficient robot path planning algorithm in static environments with danger zones. In: 2013
algorithms. IEEE Trans Intell Transp Syst 2020;23(1):572–86. http://dx.doi.org/ 17th international conference on system theory, control and computing (ICSTCC).
10.1109/TITS.2020.3014088. IEEE; 2013, p. 434–9. http://dx.doi.org/10.1109/ICSTCC.2013.6688997.
[8] Erdogan N, Erden F, Kisacikoglu M. A fast and efficient coordinated vehicle-to- [31] Zhang X, Wang Z, Lu Z. Multi-objective load dispatch for microgrid with electric
grid discharging control scheme for peak shaving in power distribution system. vehicles using modified gravitational search and particle swarm optimization
J Mod Power Syst Clean Energy 2018;6(3):555–66. http://dx.doi.org/10.1007/ algorithm. Appl Energy 2022;306:118018. http://dx.doi.org/10.1016/j.apenergy.
s40565-017-0375-z. 2021.118018.

17
K. Pan et al. International Journal of Electrical Power and Energy Systems 157 (2024) 109766

[32] Smart J, Schey S. Battery electric vehicle driving and charging behavior observed [35] Rashedi E, Nezamabadi-Pour H, Saryazdi S. GSA: a gravitational search algo-
early in the EV project. SAE Int J Altern Powertrains 2012;1(1):27–33, https: rithm. Inf Sci 2009;179(13):2232–48. http://dx.doi.org/10.1016/j.ins.2009.03.
//www.jstor.org/stable/26168962. 004.
[33] Zhenning P, Tao Y, Keying W. Decentralized coordinated dispatch for real-time [36] Bounar N, Labdai S, Boulkroune A. PSO-GSA based fuzzy sliding mode controller
optimization of massive electric vehicles considering various interests. Proc CSEE for DFIG-based wind turbine. ISA Trans 2019;85:177–88. http://dx.doi.org/10.
2019;39(12):13. http://dx.doi.org/10.13334/j.0258-8013.pcsee.172688. 1016/j.isatra.2018.10.020.
[34] Neubauer J, Wood E. The impact of range anxiety and home, workplace, [37] Tiandp S. MPSO: Modified particle swarm optimization and its applications.
and public charging infrastructure on simulated battery electric vehicle lifetime Swarm Evol Comput 2018;41:49–68. http://dx.doi.org/10.1016/j.swevo.2018.01.
utility. J Power Sources 2014;257:12–20. http://dx.doi.org/10.1016/j.jpowsour. 011.
2014.01.075.

18