Electric vehicle charging station planning method based on CSO-ELM

Abstract

The invention discloses an electric vehicle charging station planning method based on CSO-ELM. The electric vehicle charging station planning method comprises the following steps: collecting historical data of the inventory of electric vehicles; an extreme learning machine optimized based on a chicken flock algorithm is adopted to predict the electric vehicle inventory; determining charging characteristics and road characteristics of an electric vehicle user; space-time distribution prediction of the charging demand of the electric vehicle is realized by using the Latin hypercube; establishing a charging station locating and sizing model by taking the lowest sum cost of the investor construction cost, the operator maintenance cost and the electric vehicle user charging cost as a target function, and setting constraint conditions; and solving the charging station locating and sizing model through an improved particle swarm algorithm based on a single-dimensional search volume probability variation method according to a space-time distribution prediction result of the charging demand of the electric vehicle, determining the optimal station location and number of the charging stations, and completing the planning of the charging stations. According to the method, the benefits of the charging station and the EV user are comprehensively considered to construct the EV charging station locating and sizing model, the economy of the charging station is optimized, and meanwhile the satisfaction degree of the user to the charging service can be guaranteed.

Description

Technical Field

[0001] The invention relates to the technical field of site selection and capacity determination of electric vehicle charging stations, and in particular to a CSO-ELM-based electric vehicle charging station planning method. Background Art

[0002] In recent years, with the development of electric vehicle (EV) technology and the improvement of people's living standards, the number of electric vehicles (EV) has increased rapidly. The large-scale access of electric vehicles (EV) will have an important impact on the operation of the power grid, traffic economy and current political planning. As the main infrastructure for charging electric vehicles (EV), the number of EV charging stations and public charging piles is far from the preset target of "1:1 ratio of vehicles to piles" proposed by the National Energy Administration. It is expected that the number of EVs and charging piles will continue to rise rapidly in the future. The sharp increase in the demand for the construction of EV charging stations has brought new challenges to charging station operators.

[0003] On the one hand, charging station operators need to consider the economic efficiency of construction costs and the reliability of system operation to ensure that electric vehicle (EV) charging stations are highly integrated with existing infrastructure; on the other hand, they also need to take into account the spatiotemporal distribution characteristics of scattered electric vehicle (EV) users, and the built electric vehicle (EV) charging stations must meet user satisfaction. Disordered charging station planning will not only affect the economic benefits of charging station operators, but also affect user satisfaction with charging services, which is not conducive to the long-term development of the electric vehicle (EV) industry. Therefore, the optimization of the site selection and capacity of electric vehicle (EV) charging stations that takes into account both charging station revenue and user satisfaction is an urgent problem to be solved in the current research field. Summary of the invention

[0004] This paper proposes a CSO-ELM-based electric vehicle charging station planning method, which considers the impact of charging demand distribution on charging station planning while taking into account the interests of investors, operators and users, so as to achieve a more reasonable charging station planning. This method can effectively characterize the spatiotemporal distribution of electric vehicle charging demand in different functional areas and provide a basis for charging station planning.

[0005] The technical solution adopted by the present invention is:

[0006] The CSO-ELM-based electric vehicle charging station planning method includes the following steps:

[0007] Step 1: Collect historical data on electric vehicle ownership in the planned city;

[0008] Step 2: Use extreme learning machine optimized by chicken swarm algorithm to predict the number of electric vehicles;

[0009] Step 3: Determine the charging characteristics of electric vehicle users and road characteristics;

[0010] Step 4: Use Latin hypercube to achieve the spatio-temporal distribution prediction of electric vehicle charging demand;

[0011] Step 5: Taking the sum of the lowest costs of the construction cost of the investor, the maintenance cost of the operator, and the charging cost of electric vehicle users as the objective function, establish a siting and sizing model for charging stations and set up constraints;

[0012] Step 6: According to the spatio-temporal distribution prediction results of electric vehicle charging demand in Step 4, solve the siting and sizing model of the charging station in Step 5 through an improved particle swarm optimization algorithm based on the probability mutation method of single-dimensional search volume, determine the optimal site and quantity of the charging station, and complete the charging station planning.

[0013] In the said Step 2, predicting the electric vehicle ownership includes the following steps:

[0014] S2.1: In the present invention, the input data of the extreme learning machine (ELM) is the vehicle ownership value in the previous 4 years, and the output data is the vehicle ownership value in the 5th year. That is, starting from 2010, the vehicle ownership values from 2010 to 2013 are used as the first group of samples, and then the samples are taken in a rolling manner in sequence. That is, the vehicle ownership data from 2011 to 2014 is used as the second group of samples, and the subsequent sample groups follow this rule.

[0015] The regression function of the extreme learning machine (ELM) can be described as:

[0016]

[0017] In formula (1): G(·) is the activation function; ω i is the input weight of the i-th hidden layer neuron; x j is the input sample; b i is the bias of the i-th hidden layer neuron; β i is the weight between the hidden layer and the output layer, I is the number of hidden layer neurons, and m represents the number of input samples.

[0018] The ultimate goal of regression training is to minimize the error between the regression value f(x) and the corresponding actual value y j , that is, to explore the input weight ω i and the bias b i to satisfy equation (2):

[0019]

[0020] Converting the above equation into a matrix gives:

[0021]

[0022] In formula (3): β is the output weight matrix between the hidden layer and the output layer, Y is the output matrix, and H is the hidden layer output matrix, which can be expressed as:

[0023]

[0024] In formula (4): G(ω1x1 + b1) is the first hidden layer function when inputting x1; G(ω n x1 + b n ) is the nth hidden layer function when inputting x1; G(ω1x m + b1) is the first hidden layer function when inputting x m ; G(ω n x m + b n ) is the nth hidden layer function when inputting x m ; S2.2: In the Chicken Swarm Optimization (CSO), the rooster plays a dominant role during the entire foraging process. The rooster position update formula is as follows:

[0025] x i,j (t + 1) = x i,j (t)·(1 + randn(0, σ 2 )) (5);

[0026]

[0027] In the above formula: x i,j (t + 1) is the position of the jth dimension of the ith rooster at the (t + 1)th iteration; x i,j (t) represents the position coordinate of the jth dimension of the ith individual at the tth iteration. randn(0, σ 2 ) is a Gaussian distribution function with a mean of 0 and a variance of σ 2 ; σ is the scale parameter; f k is the fitness value of randomly selected rooster k; f i is the fitness value of individual i; ε is a sufficiently small positive number to prevent the denominator of from taking the value of 0; k is the number of any rooster individual, and k ≠ i; i is the ith rooster; N * is the total number of roosters;

[0028] The hen position update formula is as follows:

[0029] x i,j (t + 1) = x i,j (t) + S1·rand·(x r1,j (t) - x i,j (t)) + S2·rand·(x r2,j (t) - x i,j (t)) (7);

[0030]

[0031] S2 = exp(f r2 -f i )

[0032] Wherein, S1 is the influence factor of the partner rooster r1 on the i-th hen; S2 is the influence factor of the randomly selected rooster r2 on the i-th hen; x r1,j (t) is the j-th dimensional position of the partner rooster r1; x r2,j (t) is the j-th dimensional position of the randomly selected rooster r2; f r2 is the fitness value of the randomly selected rooster r2; r1 is the number of the mate rooster individual of the i-th hen; r2 is the number of any randomly selected rooster or hen individual, and r1 ≠ r2; rand is a random number within [0, 1].

[0033] The chick position update formula is as follows:

[0034] x i,j (t + 1) = x i,j (t) + FL·(x m,j (t) - x i,j (t)), FL ∈ [0, 2] (9);

[0035] In formula (9): x m,j (t) represents the position of the i-th chick mother at the t-th iteration; FL is the following coefficient.

[0036] S2.3: Randomly initialize the positions of the roosters in the Chicken Swarm Optimization (CSO) algorithm, and each rooster position corresponds to a set of input weights and biases of the Extreme Learning Machine (ELM).

[0037] S2.4: Take the prediction error of the Extreme Learning Machine (ELM) model corresponding to each rooster position on the training set as the fitness value.

[0038]

[0039] In the formula: f i is the fitness value of the i-th rooster; RMSE is the prediction error; N * is the total number of roosters; y i is the i-th actual value;

[0040] f(x i ) is the i-th predicted value;

[0041] S2.5: According to the update rule of the Chicken Swarm Optimization (CSO) algorithm, update the position of each rooster, that is, update the input weights and biases of the Extreme Learning Machine (ELM), that is, update the output matrix H of formula (4).

[0042] S2.6: Select the position of the rooster with the lowest fitness value (the smallest prediction error) as the current optimal solution.

[0043]

[0044] In the formula: f b is the fitness value of the b-th rooster; f1, f2, …, f N are the fitness values of N * roosters respectively; is the optimal solution within the current j-th dimension; x b,j is the position of the b-th rooster in the j-th dimension.

[0045] S2.7: Repeat steps S2.4 - S2.6 until the termination condition is met, such as the maximum number of iterations or the minimum error, and terminate the iteration when either of them is satisfied.

[0046]

[0047] In the formula: t is the number of iterations; T is the set maximum number of iterations; RMSE is the prediction error; RMSE set is the set minimum error.

[0048] S2.8: Use the optimal input weights and biases and solve the output weights of the extreme learning machine (ELM) by the least squares method.

[0049] As can be seen from Equation (3):

[0050] β = H -1 Y

[0051] In the formula: H -1 is the inverse matrix of the hidden layer output matrix H.

[0052] In step 3, the specific characteristics included are: the unit energy consumption of roads at different levels, the judgment index for EV user charging, and user travel habits;

[0053] 1) The relationship between the unit energy consumption of roads at different levels and the driving speed:

[0054]

[0055] In Equation (10): and represent the unit energy consumption of electric vehicles driving on first-class, second-class, and third-class roads; represents the driving speed of electric vehicles on section V ij during the T time period. For details, see Reference [1]: Song Yuanyuan. Research on Energy Consumption Modeling and Endurance Mileage Estimation of Pure Electric Vehicles Based on Driving Conditions [D]. Beijing Jiaotong University, 2014.

[0056] 2) EV user charging judgment indicators:

[0057] The remaining power of the vehicle at the end of a trip is as:

[0058]

[0059] In formula (11): Q o (i) represents the initial power of the vehicle; L i,j represents the road section length; represents the power consumption per unit of the road section; is the power consumed by the vehicle for a trip.

[0060] When the remaining power Q d (i) meets any condition of formula (3), a charging demand is triggered;

[0061]

[0062] In formula (12): λ is the percentage of the remaining power, and its usual value range is 0.2 - 0.35; L o→d is the distance of the next driving road section. That is, when the remaining battery power is lower than the initial set threshold or the destination cannot be reached, a charging demand is generated.

[0063] The said step 4 includes the following steps:

[0064] S4.1: Import data such as the urban road network structure, traffic flow in each time period, vehicle types and quantities, etc.;

[0065] S4.2: Randomly distribute 500 taxis and private cars in the grids of each residential area and work area, and generate the driving characteristic parameters, charging characteristic parameters and their daily travel times of each vehicle through Latin hypercube sampling. The specific sampling method is as follows:

[0066] Assume that there are K random input variables among the variables to be solved, X k is any one of the random input variables, k = 1, 2,..., K; its cumulative probability distribution function can be expressed as

[0067] Y k = F k (X k )

[0068] In the formula: Y k is the output variable; F k (·) is the relationship function between the output and the input; X k is the input variable.

[0069] Divide the sampling interval into N equally spaced and non-overlapping intervals, each with a length of 1 / N. Select the midpoint of each interval as the value of Y k and then use the inverse function method to calculate the sampling value of X k . The sampling value of X k can be expressed as:

[0070]

[0071] In the formula: X kn is the sampling value of X k ; N represents the sampling scale; is the inverse function of F k (·).

[0072] S4.3: During the simulation of the electric vehicle driving process, combine the energy consumption model per unit mileage in Equation (10) of EV users during driving to update the electric vehicle power in real time, and judge whether the charging set threshold is reached according to Equation (12); once the threshold is reached, select the nearest charging station;

[0073] S4.4: Record the charging demand quantity of electric vehicles at each node at different times to complete the prediction of the spatio-temporal distribution of charging demand. The specific steps of Step 5 include:

[0074] Take the lowest comprehensive social cost at three levels as the objective function of the charging station planning, and the charging station planning model is:

[0075] minF = C c + C w + C EV (13);

[0076] In Equation (13): F is the comprehensive social cost; C c represents the investment cost of charging station construction; C w represents the operation and maintenance cost of the charging station;

[0077] C EV represents the annual charging cost of users;

[0078] 1) The investment cost of charging station construction mainly includes the fixed investment cost of the charging station and the construction cost of charging piles. The investment and construction cost of charging piles is:

[0079]

[0080] In Equation (14): N is the number of charging stations; C fix is the fixed investment cost of the charging station; C chg is the price of a single charging pile; is the number of chargers at the i-th charging station; r0 is the discount rate; z is the operating life;

[0081] 2) The operation and maintenance costs of the charging station mainly include the equipment inspection, maintenance costs and equipment depreciation costs. Since there are many intermediate factors involved, they are converted according to the capacity ratio. The cost model is as follows:

[0082]

[0083] In formula (15): ɑ represents the proportion coefficient of labor costs; β represents the proportion coefficient of grid connection costs; T pd represents the average daily operating time of the charging pile within a day; P chg represents the power of the charging pile; T y is a parameter for one year

[0084] 3) The layout planning of the charging station will affect the transfer cost between charging stations for users, including the power consumption cost and queuing cost of users. The annual charging cost C of users EV The specific model is as follows:

[0085]

[0086] In formula (16): C ri represents the power consumption loss generated during the process of an electric vehicle user driving to the i-th charging station; C wi represents the time-consuming cost of the user in the i-th charging station.

[0087]

[0088] In formula (17): m represents the number of charging demand points to the charging station; l ik represents the shortest driving distance from the charging demand point to the charging station; v ik represents the average driving speed of the road from the demand point to the charging station; C uat represents the equivalent economic loss of the user's travel per unit time; g represents the power consumption per unit mileage; C p represents the charging electricity price;

[0089] Considering that users may not be able to charge immediately when they drive to the charging station, the present invention adds the M / G / k queuing model to the charging time-consuming cost of users;

[0090]

[0091] In formula (18): represents the number of electric vehicles going to the i-th charging station at time n; represents the average waiting time of the i-th charging station;

[0092]

[0093] In formula (19): E T 、V Trespectively represent the expectation and variance of the process of electric vehicles following a Poisson distribution with parameter λ; N is the number of charging stations; ρ = λE T The condition for Equation (19) to hold is ρ < N.

[0094] The constraint conditions of step 5 specifically include:

[0095] 1) Total demand constraint:

[0096] To ensure that the charging stations can meet the user's needs and prevent power shortages, the present invention restricts the rated capacity of all charging stations to be not less than the total demand of EV users in the target area:

[0097]

[0098] Among them:

[0099] D = βV all C cap,c (SOC ref -SOC c )(1 + ɑ) (21);

[0100] In Equation (20): D is the total daily charging demand of the target area; C cap,c is the average rated capacity of EVs; SOC ref is the state-of-charge threshold for EV charging; SOC c is the average remaining power of EVs; α is the average loss rate of EVs during charging; V all is the total traffic capacity in the planned area, that is, the total number of EVs in the planned area; β is the proportion of EVs charging every day in the traffic flow.

[0101] 2) Coverage intensity constraint:

[0102] To prevent the distance that users need to travel for charging from being too long and ensure that the charging range covers the entire planned area, the present invention sets the distance between adjacent charging stations to be no greater than 2 times the service radius of the charging station.

[0103] R s ≤ S(DT i+1 , DT i ) ≤ 2R s (22);

[0104] In Equation (22): R s is the service range of the charging station; S(DT i+1 , DT i ) is the actual distance between node i + 1 and node i.

[0105] 3) Charging station number constraint:

[0106] The satisfaction of user charging is affected by the number of charging stations in the planned area, and the number of charging stations also affects the optimal economic cost of charging station planning. The number of charging stations is related to the total charging demand in the planned area and the upper and lower limits of the charging station planning capacity. The calculation formula is shown in Equation (23):

[0107] N min ≤N≤N max (23);

[0108] Where:

[0109]

[0110] In the formula: N min is the minimum number of charging stations; N max is the maximum number of charging stations; ceil(.) is the ceiling function; S max is the maximum capacity of a single charging station; S min is the minimum capacity of a single charging station.

[0111] In step 5, the charging station location and capacity determination model includes Equations (13) to (25).

[0112] In step 6, the particle swarm algorithm has a particle size of N, and the number of candidate points for EV charging station construction is D; the initial position of the particle is X n ={x n1 ,x n2 ,…,x nD}, n = 1, 2, …, N, where: x ni represents the construction situation of candidate point i of particle n for EV, x ni =1 means building a station at this candidate point, x ni =0 means not building a station at this candidate point. The initial velocity of the particle is: V n ={v n1 ,v n2 ,…,v nD}, v n1 ,v n2 ,…,v nD are the initial velocities of particles 1, 2, …, D respectively; P best,n is the historical best point searched by a single particle X n , G best is the global best point searched by all particles. The update formulas for the velocities and positions of all particles are:

[0113] V n,k+1 =ωV n,k +c1r1(P best,n -X n )+c2r2(G best -Xn ) (26);

[0114] X n,k+1 = X n,k + V n,k+1 ) (27);

[0115] Where: V n,k+1 is the velocity of the nth particle at the (k + 1)th iteration; V n,k is the velocity of the nth particle at the kth iteration; X n,k+1 is the position of the nth particle at the (k + 1)th iteration; X n,k is the position of the nth particle at the kth iteration; ω is the inertia weight; c1, c2 are the particle learning coefficients; r1, r2 are the mean random perturbations. In order to simulate the tiny perturbations caused by nature to increase the diversity of particles. The inertia weight is updated in a linearly decreasing form as the number of iterations increases. The update formula is as follows:

[0116]

[0117] In formula (19): ω is the inertia weight; ω max is the initial weight of the particle; ω min is the final weight of the particle; I iter,max is the maximum number of iterations; i iter is the current number of iterations.

[0118] Update rule of the learning coefficient: In the initial stage of iteration, in order for the particles to search the entire space, a larger self-learning coefficient c1 of the particle and a smaller social learning coefficient c2 are set. As the number of iterations increases, in order to make the flight target of the particles more inclined to the global optimal solution. A larger social learning coefficient c2 and a smaller self-learning coefficient c1 are set. The update formula of the learning coefficient is as follows:

[0119]

[0120] Where: c1 is the self-learning coefficient of the particle; c2 is the social learning coefficient of the particle. c 1e , c 1f , c 2e , c 2f are all constants, representing the initial value and the final value of c1 and c2 respectively. For most benchmarks, when setting c 1e , c 1f , c 2e , c 2f to be 2.5, 0.5, 0.5, 2.5 respectively, the optimization effect is better.

[0121] During the process of searching for the optimal solution, the search space of the i-th dimension (i = 1, 2, ..., I) of the particle is divided into several intervals, and the search quantity limit value of each interval is set to A max , that is, each interval can have at most A max particles for search; when the position of the i-th dimension of the particle x is about to be updated to a certain interval, if there are already A max particles searching in this interval at this time, the position of the i-th dimension of the particle x will no longer be updated to this interval, but will be randomly initialized to other intervals according to the probability of the single-dimension search quantity.

[0122] The mutation operation in the genetic algorithm changes the genetic genes on the chromosome by random selection to generate new individuals. According to this mutation operation, the present invention assumes that the search quantity of each interval in the i-th dimension of the particle should be the same, that is, for any interval, the greater the historical search quantity of this interval, the smaller the probability that the particle will enter this interval during mutation. Now the i-th dimension is evenly divided into J intervals, i = 1, 2, ..., I; when there is a particle searching in the j-th interval of the i-th dimension, j = 1, 2, …, J, the search quantity update formula of the interval is as follows:

[0123] A i,j = A i,j + 1 (31);

[0124] In formula (31): A i,j is the number of search times of the particle in the j-th interval of the i-th dimension;

[0125] The total search quantity of the particle in the i-th dimension interval is:

[0126]

[0127] In formula (32): J is the number of intervals of the dimension; sum i (A) is the total search quantity of the particle in the i-th dimension interval;

[0128] For any particle x, when the position of its i-th dimension chooses to mutate, the probabilities of mutating to each interval are respectively:

[0129] P i = [A iJ , A iJ-1 , …, A i1 / sum i (A) (33);

[0130] In formula (33): P i is the probability of the particle mutating to each interval when it reaches the i-th dimension;

[0131] That is, the mutation probability of each interval in the i-th dimension changes with the update of the position of each particle in this dimension, and it is closely related to the historical search volume of this interval.

[0132] A method for planning an electric vehicle charging station based on CSO-ELM according to the present invention has the following technical effects:

[0133] 1) The method for planning an electric vehicle charging station of the present invention comprehensively considers the interests of both the charging station and EV users to construct an EV charging station site selection and capacity determination model, which can optimize the economy of the charging station while ensuring the satisfaction of users with charging services.

[0134] 2) The method for planning an electric vehicle charging station of the present invention considers the differences of different charging piles, and studies the capacity configuration problem accurately to the charging pile, which is more in line with the engineering practice compared with the overall capacity configuration method of the charging station.

[0135] 3) The present invention proposes a probability mutation method based on the single-dimensional search volume to improve the particle swarm optimization algorithm. This mutation method can increase the particle diversity, avoid the particles falling into the local optimum, and accelerate the algorithm convergence speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0136] The present invention will be further described below in conjunction with the drawings and embodiments:

[0137] Figure 1 is a schematic flow chart of a method for planning an electric vehicle charging station based on CSO-ELM;

[0138] Figure 2 is a flow chart for predicting the electric vehicle ownership in the planned city;

[0139] Figure 3 is an overall flow chart for site selection and capacity determination of an EV charging station.

[0140] Figure 4 is the training result of vehicle ownership.

[0141] Figure 5 is a comparison chart of vehicle ownership data and model prediction values.

[0142] Figure 6 is a prediction curve of electric vehicle ownership.

[0143] Figure 7 is a route planning map of a certain urban area.

[0144] Figure 8 is the spatial layout of the charging station. DETAILED DESCRIPTION OF THE INVENTION

[0145] Electric vehicle charging station planning method based on CSO-ELM. This method introduces the extreme learning machine based on chicken swarm optimization (CSO-ELM) to predict the data of the future number of electric vehicles in the planned area. The chicken swarm algorithm is used to optimize the hidden layer parameters of the extreme learning machine, and then the historical electric vehicle ownership data is substituted to train the regression equation, and finally the prediction of the future number of electric vehicles is realized. Then the Latin hypercube method is used to predict the spatio-temporal distribution of electric vehicles. On this basis, considering the interests of investors, operators, and EV users, an improved particle swarm algorithm based on the probability mutation method of single-dimensional search volume is proposed to solve the EV charging station location and capacity determination model. This method can effectively characterize the spatio-temporal distribution law of electric vehicle charging demand in different functional areas and provide a basis for charging station planning. The optimization model proposed by the present invention has the advantages of good economy and high user satisfaction. At the same time, the improved particle swarm algorithm shows the characteristics of fast optimization speed and strong adaptability in the calculation process, has good practicability, and can provide a reference for actual charging station planning.

[0146] Verification example:

[0147] Select the electric vehicle ownership in this area from 2010 to 2019 as the training sample, set the number of ELM hidden layer nodes to 5, and the parameters of the chicken swarm optimization algorithm are: the maximum number of iterations is 100, the population size is 50, the proportion of roosters is 0.2, the proportion of hens is 0.6, among which the proportion of mother hens is 0.5, and the proportion of chicks is 0.2. Use the vehicle ownership data from 2010 to 2019 to train the model, and the training results are as Figure 4 shown. It can be seen from Figure 4 that the training fitting result of this model is good. Next, compare the real vehicle ownership data from 2020 to 2024 with the model prediction values, and the test results are as Figure 5 shown. The relative error and absolute error of the model are both less than 0.1, indicating that the model can well predict the trend of vehicle ownership. Based on the road traffic flow in the selected planning area and the number of electric vehicles predicted to reach in the future data, the electric vehicle flow on each road in the future planning area is obtained by proportional expansion.

[0148] Select the urban route planning map of a certain area as Figure 7 shown, which contains a total of 31 traffic nodes, 3 main road levels, and is divided into four typical plans: residential area, working area, commercial area, and public service area. The maximum number of vehicles with charging demand in this area is set to 500. The relevant parameter settings involved in formulas (14)-(22) are shown in Table 1.

[0149] Table 1 Specific parameter settings

[0150]

[0151] It is estimated that 4 - 12 charging stations will be built in this area. The annual costs of charging stations, electric vehicle users and the power grid are shown in Table 2. Research and analysis show that with the continuous expansion of the scale of charging infrastructure, both the total annual infrastructure investment and the site operation and maintenance costs show an obvious linear growth trend; however, this growth trend effectively reduces the travel migration cost and charging waiting time cost of users. Through model calculation, it is found that when the deployment of charging piles reaches the optimal scale critical point, that is, when the number of charging stations is 7, the minimum social overall construction cost can be achieved.

[0152] Table 2 Comprehensive Social Costs of Charging Stations

[0153]

[0154] When the deployment of charging pile facilities reaches 7, the charging pile configuration plan, operation and maintenance costs and user loss costs of each site are shown in Table 3. Figure 8 As for the spatial layout of charging stations, the stations are evenly distributed, and the station spacing is set strictly in accordance with the planning constraints, effectively avoiding the problem of resource redundancy caused by over - dense stations. Specifically analyze the layout characteristics of each site: Site G is located at the intersection of the commercial business district and the employment concentration area, so the largest - scale charging facilities are configured; The functional attributes of the areas where Sites A and B are located are relatively single, but the traffic flow is relatively stable, making the user waiting cost at a relatively low level; Although Sites C and D are near the regional boundary, the traffic demand of the work areas they serve has regular characteristics; Sites E and F are located around public service facilities. Although the daily service flow is low, the user waiting cost may increase during specific peak hours.

[0155] Table 3 Planning Costs of Each Site

[0156]

Claims

1. A method for planning electric vehicle charging stations based on CSO-ELM, characterized in that The following steps are involved: Step 1: Collect historical data on electric vehicle ownership; Step 2: Use extreme learning machine optimized by chicken swarm algorithm to predict the number of electric vehicles; Step 3: Determine the charging characteristics of electric vehicle users and road characteristics; Step 4: Use Latin hypercube to predict the spatiotemporal distribution of electric vehicle charging demand; Step 5: Taking the minimum cost of the sum of the construction cost of the investor, the maintenance cost of the operator, and the charging cost of the electric vehicle user as the objective function, establish a charging station site selection and capacity determination model and set constraints; Step 6: Based on the spatiotemporal distribution prediction results of electric vehicle charging demand in step 4, the charging station location and capacity determination model in step 5 is solved by using an improved particle swarm algorithm based on the single-dimensional search volume probability mutation method to determine the optimal location and number of charging stations and complete the charging station planning.

2. The method for planning an electric vehicle charging station based on CSO-ELM according to claim 1, wherein: In step 2, predicting the number of electric vehicles includes the following steps: S2.1: The input data of the extreme learning machine ELM is the value of the number of cars in the previous four years, and the output data is the value of the number of cars in the fifth year; the regression function of the extreme learning machine ELM is described as: In formula (1): G(·) is the activation function; ω i is the input weight of the i-th hidden layer neuron; x j is the input sample; b i is the bias of the i-th hidden layer neuron; β i is the weight between the hidden layer and the output layer, I is the number of hidden layer neurons, and m represents the number of input samples; The ultimate goal of regression training is to minimize the error between the regression value f(x) and the corresponding actual value y j That is, to explore the input weights ω i and the bias b i to satisfy equation (2): Converting the above equation into a matrix: Hβ=Y (3); In formula (3), β is the output weight matrix between the hidden layer and the output layer, Y is the output matrix, and H is the hidden layer output matrix, which can be expressed as: In formula (4): G(ω1x1 + b1) is the first hidden layer function when the input is x1; G(ω n x1 + b n ) is the nth hidden layer function when the input is x1; G(ω1x m + b1) is the first hidden layer function when the input is x m ; G(ω n x m + b n ) is the nth hidden layer function when the input is x m ; S2.2: In the chicken swarm optimization algorithm CSO, the rooster plays a dominant role throughout the foraging process, and the rooster position update formula is as follows: x i，j (t + 1)= x i，j (t)·(1 + randn(0, σ 2 )) (5); Upper middle: x i，j (t + 1) is the position of the i-th rooster in the j-th dimension at the (t + 1)-th iteration; x i，j (t) represents the position coordinate of the i-th individual in the j-th dimension at the t-th iteration, randn(0, σ 2 ) is a Gaussian distribution function with a mean of 0 and a variance of σ 2 ; σ is the scale parameter; f k is the fitness value of randomly selected rooster k; f i is the fitness value of individual i; ε is a sufficiently small positive number to prevent the denominator from taking a value of 0; k is the number of any rooster individual, and k ≠ i; i is the i-th rooster; N * is the total number of roosters; The formula for updating the hen's position is as follows: x i，j (t + 1) = x i，j (t) + S1·rand·(x r1，j (t) - x i，j (t)) + S2·rand·(x r2，j (t) - x i，j (t)) (7); Among them, S1 is the influence factor of the partner rooster r1 on the i-th hen; S2 is the influence factor of the randomly selected rooster r2 on the i-th hen; x r1，j (t) is the j-th dimensional position of the partner rooster r1; x r2，j (t) is the j-th dimensional position of the randomly selected rooster r2; f r2 is the fitness value of the randomly selected rooster r2; r1 is the number of the mate rooster individual of the i-th hen; r2 is the number of any randomly selected rooster or hen individual, and r1 ≠ r2; rand is a random number within [0, 1]; The chicken position update formula is as follows: x i，j (t + 1)=x i，j (t)+FL·(x m，j (t)-x i，j (t)), FL ∈ [0, 2] (9); In Equation (9): x m，j (t) represents the position of the i-th chicken mother at the t-th iteration; FL is the following coefficient.

3. The method for planning an electric vehicle charging station based on CSO-ELM according to claim 2, wherein: It also includes S2.3: randomly initializing the positions of the roosters in the chicken swarm algorithm CSO, each rooster position corresponds to a set of input weights and biases of the extreme learning machine ELM; S2.4: The prediction error of the extreme learning machine ELM model corresponding to each rooster position on the training set is used as the fitness value; where: f i is the fitness value of the i-th rooster; RMSE is the prediction error; N * is the total number of roosters; y i is the i-th actual value; f(x i ) is the i-th predicted value; S2.5: According to the update rule of the chicken swarm algorithm CSO, update the position of each rooster, that is, update the input weight and bias of the extreme learning machine (ELM), that is, update the output matrix H of formula (4); S2.6: Select the rooster position with the lowest fitness value as the current optimal solution; where: f b is the fitness value of the b-th rooster; f1, f2, …, f N are respectively the fitness values of N * roosters; is the optimal solution within the current j-th dimension; x b，j is the position of the b-th rooster in the j-th dimension; S2.7: Repeat steps S2.4-S2.6 until the termination condition is met, such as the maximum number of iterations or the minimum error, and the iteration is terminated when one of the two conditions is met; Where: t is the number of iterations; T is the set maximum number of iterations; RMSE is the prediction error; RMSE set is the set minimum error; S2.8: Using the optimal input weights and biases, the least squares method is used to solve the output weights of the extreme learning machine (ELM); From formula (3), we can see that: β = H -1 Y Where: H -1 is the inverse matrix of the hidden layer output matrix H.

4. The method for planning an electric vehicle charging station based on CSO-ELM according to claim 1, wherein: In step 3, the relationship between unit energy consumption and driving speed of roads of different grades is: In formula (10): and represent the unit energy consumption of the electric vehicle when driving on first-class, second-class, and third-class roads; represents the driving speed of the electric vehicle on section V during period T.

5. The method for planning an electric vehicle charging station based on CSO-ELM according to claim 4, characterized in that: In step 3, the EV user charging judgment index is: The remaining power of the vehicle at the end of a trip is : In Equation (11): Q(i) represents the initial battery level of the vehicle; L ， represents the length of the road segment; represents the power consumption per unit length of the road segment; is the power consumption for the vehicle to complete one trip; When the remaining power Q(i) satisfies any condition of formula (3), the charging demand is triggered; In formula (12): λ is the percentage of remaining battery power; L → is the distance of the next driving section; that is, a charging requirement is generated when the remaining battery power is lower than the initial set threshold or the destination cannot be reached.

6. The method for planning an electric vehicle charging station based on CSO-ELM according to claim 1, wherein: The step 4 comprises the following steps: S4.1: Import data on urban road network structure, traffic flow at different time periods, vehicle types and quantities, etc. S4.2: Randomly distribute taxis and private cars in each residential area and work area grid, and generate the driving characteristic parameters, charging characteristic parameters and the number of trips per day of each vehicle through Latin hypercube sampling; S4.3: During the simulation of the electric vehicle driving process, the electric vehicle power is updated in real time according to the energy consumption model per unit mileage in Equation (10) of EV users during driving, and it is judged whether the charging set threshold is reached according to Equation (12); once the threshold is reached, the nearest charging station is selected for charging. S4.4: Record the charging demand quantities of electric vehicles at each node in different time periods to complete the prediction of the spatio-temporal distribution of charging demands.

7. The method for planning an electric vehicle charging station based on CSO-ELM according to claim 6, wherein: In S4.2, the specific sampling method is as follows: Suppose there are K random input variables among the variables to be solved, and X k is any one of the random input variables, where k = 1, 2, …, K; Its cumulative distribution function can be expressed as Y k = F k (X k ) Where: Y k is the output variable; F k (·) is the relationship function between the output and the input; X k is the input variable; Divide the sampling interval into N equally spaced and non-overlapping intervals, each with a length of 1 / N, and select the midpoint of each interval as the value of Y k Then use the inverse function method to calculate the sampling value of X k ; The sampling value of X k can be expressed as: Where: X kn is the sampled value of X k ; N represents the sampling scale; is the inverse function of F k (·).

8. The method for planning an electric vehicle charging station based on CSO-ELM according to claim 1, wherein: The specific steps of Step 5 include: Taking the lowest comprehensive social cost at three levels as the objective function for charging station planning, the charging station planning model is: minF = C c + C w + C EV (13); In formula (13): F is the comprehensive social cost; C c represents the construction investment cost of the charging station; C w represents the operation and maintenance cost of the charging station; C EV represents the annual charging cost of users; 1) The investment cost for charging station construction includes the fixed investment cost of the charging station and the construction cost of charging piles. The investment and construction cost of charging piles is: In formula (14): N is the number of charging stations; C fx is the fixed investment cost of the charging station; C cg is the price of a single charging pile; is the number of chargers at the i-th charging station; r0 is the discount rate; z is the operation period; 2) The operation and maintenance cost of the charging station includes the equipment inspection, maintenance cost and equipment depreciation cost of the charging station. Since there are many intermediate factors involved, it is converted according to the capacity ratio. Its cost model is: In formula (15): α represents the coefficient of the labor cost ratio; β represents the coefficient of the grid connection cost ratio; T p represents the daily average operating time of the charging pile within a day; P cg represents the power of the charging pile; T y is the parameter for one year 3) The layout planning of the charging stations will affect the transfer costs between charging stations for users, including the power consumption cost and queuing cost of users. The annual charging cost C of users EV The specific model is as follows: In formula (16): C r represents the power consumption loss generated during the process of an electric vehicle user driving to the i-th charging station; C w represents the time-consuming cost of the user in the i-th charging station. In formula (17): m represents the number of charging demand points going to the charging station; l k represents the shortest driving distance from the charging demand point to the charging station; v k represents the average driving speed of the road from the demand point to the charging station; C ua represents the equivalent economic loss of the user's travel per unit time; g represents the power consumption per unit mileage; C p represents the charging electricity price; Considering that users may not charge immediately when they drive to the charging station, the M / G / k queuing model is added to the charging time cost of users. In Equation (18): represents the number of electric vehicles going to the i-th charging station at time n; represents the average waiting time at the i-th charging station; In Equation (19): E and V respectively represent the expectation and variance of the electric vehicle process following a Poisson distribution with parameter λ; N is the number of charging stations; ρ = λE T , and the condition for Equation (19) to hold is ρ < N.

9. The method for planning an electric vehicle charging station based on CSO-ELM according to claim 8, wherein: The constraint conditions of Step 5 specifically include: 1) Total demand constraint: Restrict that the rated capacity of all charging stations is not less than the total demand of EV users in the target area: Where: D = βV all C cap，c (SOC ref - SOC c )(1 + α)(21); In Equation (20): D is the total daily charging demand of the target area; C cap，c is the average rated capacity of the EV; SOC ref is the state-of-charge threshold for EV charging; SOC c is the average remaining power of the EV; α is the average loss rate of the EV during charging; V all is the total traffic capacity in the planning area, i.e., the total number of EVs in the planning area; β is the proportion of EVs charging daily in the traffic flow; 2) Coverage intensity constraint: Set the distance between two adjacent charging stations to be no greater than 2 times the service radius of the charging station. R s ≤S(DT i+1 ,DT i )≤2R s (22); In formula (22): R s is the service range of the charging station; S(DT i+1 , DT i ) is the actual distance between node i + 1 and node i; 3) Charging station quantity constraint: The quantity of charging stations is related to the total charging demand in the planning area and the upper and lower limits of the charging station planning capacity. The calculation formula is shown in Equation (23): N min ≤N≤N max (23); Where: Where: N min is the minimum number of charging stations; N max is the maximum number of charging stations; ceil(.) is the ceiling function; S max is the maximum capacity of a single charging station; S min is the minimum capacity of a single charging station.

10. The method for planning an electric vehicle charging station based on CSO-ELM according to claim 8, wherein: In step 6, the particle swarm size of the particle swarm algorithm is N, and the number of candidate points for EV charging station construction is D; the initial position of the particle is X n ={x n1 , x n2 , …, x nD}, n = 1, 2, …, N, where: x n represents the construction situation of candidate point i of particle n, x ni = 1 means building a station at this candidate point, x n = 0 means not building a station at this candidate point; the initial velocity of the particle is V n ={v n1 , v n2 , …, v nD}, v n1 , v n2 , …, v nD are the initial velocities of particles 1, 2, …, D respectively; P best，n is the historical best point searched by a single particle X n , and G bes is the global best point searched by all particles; the velocity and position update formulas for all particles are: V n，k+1 = ωV n，k + c1r1(P best，n - X n ) + c2r2(G best - X n ) (26); X n，k+1 = X n，k + V n，k+1 (27); Where: V n，k+1 is the velocity of the nth particle at the (k + 1)th iteration; V n，k is the velocity of the nth particle at the kth iteration; X n，k+1 is the position of the nth particle at the (k + 1)th iteration; X n，k is the position of the nth particle at the kth iteration; ω is the inertia weight; c1 and c2 are the particle learning coefficients; r1 and r2 are the mean random perturbations, in order to simulate the tiny perturbations caused by nature to increase the diversity of particles; the inertia weight is updated in a linearly decreasing form as the number of iterations increases, and the update formula is as follows: In Equation (19): ω is the inertia weight; ω ax is the initial weight of the particle; ω n is the final weight of the particle; I er，ax is the maximum number of iterations; i er is the current number of iterations; Update rule of learning coefficient: In the initial stage of iteration, in order for the particles to search in the entire space, set a larger self-learning coefficient c1 of the particles and a smaller social learning coefficient c2; as the number of iterations increases, in order to make the flight target of the particles more inclined to the global optimal solution; set a larger social learning coefficient c2 and a smaller self-learning coefficient c1; the learning coefficient update formula is shown as follows: Where: c1 is the self-learning coefficient of the particle; c2 is the social learning coefficient of the particle; c 1e ,c 1f ,c 2e ,c 2f are all constants, representing the initial value and the final value of c1 and c2 respectively; During the process of searching for the optimal solution, the search space of the i-th dimension (i = 1, 2,..., I) of the particle is divided into several intervals, and the search quantity limit value of each interval is set to A ax , that is, each interval can have at most A ax particles for searching; when the i-th dimensional position of particle x is about to be updated to a certain interval, if there are already A ax particles searching in this interval at this time, the position of the i-th dimension of particle x will no longer be updated to this interval, but will be randomly initialized to other intervals according to the probability of the single-dimensional search quantity; Suppose the search quantity of each interval in the i-th dimension of the particle should be the same, that is, for any interval, the larger the historical search quantity of this interval, the smaller the probability that the particle enters this interval during mutation; now divide the i-th dimension into J intervals, i = 1, 2,..., I; when there is a particle searching in the j-th interval of the i-th dimension, j = 1, 2,..., J, the search quantity update formula of the interval is as follows: A i，j = A i，j + 1 (31); In formula (31): A i，j is the search times of the particle in the j-th interval of the i-th dimension; The total search quantity of the particle in the i-th dimension interval is: In formula (32): J is the number of intervals of dimensions; sum i (A) is the total search amount of the particle in the i-th dimension interval; For any particle x, when the position of its i-th dimension selects mutation, the probabilities of mutating into each interval are respectively: P i = [A iJ , A iJ-1 , …, A i1 / sum i (A) (33); In formula (33): P i is the probability of each interval when the particle mutates to the i-th dimension; That is, the mutation probability of each interval in the i-th dimension changes with the update of the position of each particle in this dimension, and it is closely related to the historical search quantity of this interval.

Images