A method and system for predicting energy of urban electric taxis

Abstract

The present invention discloses a method and system for predicting energy of urban electric taxis. The method trains electric taxis according to the initialization model requirements of a charging station service provider to obtain a daily charging behavior model of the electric taxis, and transmits the generated daily charging behavior model parameters to the charging station; the charging station can perform aggregation according to the obtained daily charging behavior model parameters of the electric taxis, optimize the local model of the charging station according to the aggregated parameters, and transmit the optimized local model parameters to the charging station service provider for full aggregation to obtain a full aggregation prediction model. The full aggregation prediction model is used to predict the energy of the electric taxis, which solves the problem of insufficient energy prediction accuracy due to insufficient training set data in machine learning. The method of the present invention can achieve energy supply and demand balance and low-cost power transmission, realize timely energy storage during power troughs, and meet normal charging requests of electric taxis.

Description

Technical Field

[0001] The invention belongs to the technical field of electric vehicle management, and in particular relates to an energy prediction method and system for urban electric taxis. Background Art

[0002] Power rationing, shutdown, and peak-shifting electricity consumption reflect the severe power supply situation. The electrification of urban transportation has started with taxis. Energy regulation between charging station suppliers, charging stations, and electric taxis will be an effective way to solve the problem of urban power optimization. Since electric taxis have complex charging behaviors and large charging demands, they impose a serious energy burden on the power system, which will directly lead to other equipment being underpowered.

[0003] Existing energy predictions based on machine learning are based on charging stations, and do not consider the energy consumption end, i.e., electric taxis. The common point of these methods is that they do not form a systematic relationship between electric taxis, charging stations, and charging station service providers, and do not involve the issue of privacy protection. Due to the small amount of data from charging stations in cities, the accuracy of machine learning is seriously insufficient. There is an urgent need for a systematic electric taxi energy prediction method, which plays an important role in regulating electricity for relevant departments and also accelerates the development of smart cities in our country. Summary of the invention

[0004] The purpose of the present invention is to provide an energy prediction method and system for urban electric taxis to overcome the shortcomings of the prior art, and to solve the problems of insufficient training samples, lack of privacy protection, and weak systematicity in the prior art. The energy prediction of the present invention is highly accurate and can provide a reliable basis for energy regulation of relevant departments.

[0005] A method for predicting energy of urban electric taxis comprises the following steps:

[0006] S1, training the electric taxi according to the initialization model requirements of the charging station service provider, obtaining a daily charging behavior model of the electric taxi, and transmitting the generated daily charging behavior model parameters to the charging station;

[0007] S3, the charging station can aggregate the daily charging behavior model parameters of the electric taxis, optimize the local model of the charging station according to the aggregated parameters, transmit the optimized local model parameters to the charging station service provider for full aggregation, obtain the full aggregation prediction model, and use the full aggregation prediction model to predict the energy of the electric taxis.

[0008] Furthermore, the electric taxi module adopts the local model of the electric taxi, the communication network module is used for data communication transmission, and the charging station module is used to simulate charging at the charging station.

[0009] Furthermore, the electric taxi model makes local predictions based on the relevant data of the electric taxi's own charging, and uploads the weight change value to the charging station model through the communication network.

[0010] Furthermore, the local model corresponding to the electric taxi module simulates an operating taxi that is driven solely by electricity.

[0011] Furthermore, the parameters of the daily charging behavior model are obtained through the charging data collection of electric taxi owners.

[0012] Furthermore, the electric taxi module locally trains the daily charging behavior model using a machine learning training method.

[0013] Furthermore, let the input matrix be X i ={x1,x2,…,x p}, x p For the charging attribute set of the pth electric taxi, a local network model can be established:

[0014] The batch gradient descent method can be used to obtain the weight update method of the electric taxi:

[0015]

[0016] Gradient information for training the local network model of an electric taxi charging at a charging station will be transmitted to the charging station; the weight vector of the charging station provider is ω csp ; Weight vector of charging station Weight vector of electric taxi i and j represent the number of electric taxis and charging stations, respectively.

[0017] Furthermore, charging probability statistics and Gaussian distribution probability values ​​are used as attention to form aggregation weights.

[0018] Furthermore, the loss function and weight update method of the charging station are:

[0019]

[0020]

[0021] An urban electric taxi energy prediction system includes a first-layer federated learning system and a second-layer federated learning system;

[0022] The first-level federated learning system includes an electric taxi module, a communication network module, and a local charging station module. The electric taxi module is trained locally based on the local charging station module to obtain a daily charging behavior model.

[0023] The second-layer federated learning system includes a charging station module, an Internet module, and a charging station provider module. The charging station module is used to obtain the daily charging behavior model parameters of the electric taxi module and then aggregate them, and optimize the local model of the charging station based on the aggregated parameters. The optimized local model parameters are sent to the charging station provider module through the Internet module. The charging station provider module performs the final global aggregation to obtain a fully aggregated prediction model.

[0024] Compared with the prior art, the present invention has the following beneficial technical effects:

[0025] The present invention provides an urban electric taxi energy prediction method. The electric taxi is trained according to the initialization model requirements of a charging station service provider to obtain a daily charging behavior model of the electric taxi, and the generated daily charging behavior model parameters are transmitted to the charging station; the charging station can perform aggregation according to the obtained daily charging behavior model parameters of the electric taxi, optimize the local model of the charging station according to the aggregated parameters, and transmit the optimized local model parameters to the charging station service provider for full aggregation to obtain a full aggregation prediction model. The full aggregation prediction model is used to predict the energy of the electric taxi, which solves the problem of insufficient energy prediction accuracy due to insufficient training set data in machine learning. The method of the present invention can achieve energy supply and demand balance and low-cost power transmission, realize timely energy storage during power trough periods, and meet normal charging requests of electric taxis.

[0026] Furthermore, for charging station providers in the second layer of federated learning, the goal is to minimize the loss function of the local model and solve the quality imbalance problem when charging stations aggregate local models. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a specific system module connection and principle diagram in an embodiment of the present invention;

[0028] Figure 2 This is an application example diagram in the embodiment of the present invention.

[0029] Figure 1 :1. First-layer federated learning system, 11. Communication network, 12. Network model (GRU), 13. Electric taxi, 2. Second-layer federated learning system, 21. Charging station provider, 22. Internet, 23. Charging station. DETAILED DESCRIPTION

[0030] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0031] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0032] A method for predicting energy of urban electric taxis comprises the following steps:

[0033] S1, training the electric taxi according to the initialization model requirements of the charging station service provider, obtaining a daily charging behavior model of the electric taxi, and transmitting the generated daily charging behavior model parameters to the charging station;

[0034] S3, the charging station can aggregate the daily charging behavior model parameters of the electric taxis, optimize the local model of the charging station according to the aggregated parameters, transmit the optimized local model parameters to the charging station service provider for full aggregation, obtain the full aggregation prediction model, and use the full aggregation prediction model to predict the energy of the electric taxis.

[0035] The communication network module of the present application adopts communication network communication transmission.

[0036] The electric taxi module adopts the local model of electric taxi, the communication network module is used for data communication transmission, and the charging station module is used to simulate charging at the charging station.

[0037] The electric taxi model makes local predictions based on the relevant data of the electric taxi’s own charging, and uploads the weight change values ​​to the charging station model through the communication network.

[0038] The daily charging behavior model parameters are derived from the charging data of electric taxi owners.

[0039] The electric taxi module locally trains the daily charging behavior model using a machine learning training method.

[0040] The present invention is directed to a method for predicting energy of urban electric taxis, which mainly includes:

[0041] Construct a two-layer federated learning of electric taxi-charging station and charging station-charging station provider. Here, the charging station provider sends the initialization model to the electric taxi, and the electric taxi trains its own daily charging behavior model. The generated daily charging behavior model parameters are sent to the charging station. After aggregation, the charging station optimizes its own local model and sends the optimized gradient information to the charging station provider. The charging station provider performs the final global aggregation to obtain the fully aggregated prediction model, and uses the fully aggregated prediction model to predict the energy of urban electric taxis.

[0042] like Figure 1 As shown, an urban electric taxi energy prediction system includes a first-layer federated learning system 1 and a second-layer federated learning system 2;

[0043] The first-layer federated learning system 1 includes an electric taxi module, a communication network module, and a local charging station module. The local training end of the first-layer federated learning is the electric taxi module, and the model aggregation end is the charging station module. The electric taxi module trains the daily charging behavior model locally according to the local charging station module, and sends the daily charging behavior model parameters generated by the training to the charging station module through the communication network module;

[0044] The second-layer federated learning system includes a charging station module, an Internet module, and a charging station provider module. The local training end of the second-layer federated learning is the charging station module, and the aggregation end is the charging station provider module. The charging station module aggregates after obtaining the daily charging behavior model parameters of the electric taxi module, and optimizes the local model of the charging station based on the aggregated parameters. The optimized local model parameters are sent to the charging station provider module through the Internet module. The charging station provider module performs the final global aggregation to obtain a fully aggregated prediction model.

[0045] The charging station model is used to obtain the relevant parameters of the daily charging behavior model of electric taxis and then aggregate them, and send the relevant change information to the charging station supplier module through the Internet module, and the charging station supplier performs the final global aggregation.

[0046] The local model corresponding to the electric taxi module simulates an operating taxi that is driven solely by electricity.

[0047] In the fully aggregated prediction model, the weight vector of the charging station provider is ω csp ; Weight vector of charging station Weight vector of electric taxi i and j represent the number of electric taxis and charging stations.

[0048] Let the input matrix be X i ={x1,x2,…,x p}, x p For the charging attribute set of the pth electric taxi, a local network model can be established:

[0049] Here, the matrix The size is h×p, b i is the network bias, h is the number of network layers, f i (·) represents the RELU activation function.

[0050] The loss function of each electric taxi is:

[0051]

[0052] The batch gradient descent method can be used to obtain the weight update method of the electric taxi:

[0053]

[0054] Gradient information for training the local network model of an electric taxi charging at a charging station It will be transmitted to the charging station. In the first layer of federated learning, the charging station acts as a model aggregator, and the electric taxi outputs the energy demand model for a certain period.

[0055] Because charging stations pay different attention to each electric taxi, different electric taxi parameters are selected for aggregation at different charging stations, which will further solve the problem of unbalanced model quality.

[0056] The charging probability statistics and Gaussian distribution probability values ​​are selected as attention to form the aggregation weights.

[0057] Assume E j The probability value of selecting a charging station for an electric vehicle and obtaining the maximum probability of selecting the jth charging station is:

[0058]

[0059]

[0060] Where P xk Represents the number of charging times.

[0061] The probability of an electric taxi choosing a charging station in the region also follows a two-dimensional Gaussian distribution. The parameters generated by the two-dimensional Gaussian distribution model should be introduced into the attention mechanism of statistical probability:

[0062]

[0063] Where: A is a constant, represents the probability of selecting the jth charging station in a certain area, L(k) represents the activity point of the electric taxi near the kth charging station before charging, and L(j) is the location of the jth charging station.

[0064] In summary, the loss function and weight update method of the charging station are:

[0065]

[0066]

[0067] For charging station providers in the second layer of federated learning, the goal is to minimize the loss function of the local model (charging station), which can solve the problem of quality imbalance when charging stations aggregate local models.

[0068] Given the geographical characteristics of the distribution of charging station data, charging station service providers focus more on the service areas of charging stations and the correlation between charging stations.

[0069] When performing the aggregation of the second-layer federated learning system, the geographical fence information, namely the Manhattan distance function, is introduced:

[0070]

[0071] here Represents the central location of a certain area. Calculate the Manhattan ratio of all charging stations and normalize them to get:

[0072]

[0073] Where:

[0074] Another aspect is that there is a relationship of service competition and speed competition among charging stations. Suppose the correlation coefficient between the jth charging station and its closest charging station.

[0075]

[0076] In the formula represents the combination of the frequency of charging vehicles and waiting time within a certain period of the charging station. At this time, the loss function of our charging station provider can be expressed as:

[0077]

[0078] Under the weighted method of the multi-attention mechanism, the model of charging station provider weight update can be obtained:

[0079]

[0080] In the two-layer federated learning, the charging station provider absorbs the gradient information of the charging station until the charging station provider model parameters are aggregated, and the global loss function L g (ω csp ) takes the minimum value, and the termination condition is that the prediction error converges to a certain value s or reaches a certain number of iterations T.

[0081] The present invention discloses an energy prediction method for urban electric taxis. The electric taxis are trained according to the initialization model requirements of the charging station service provider, thereby solving the problem of insufficient energy prediction accuracy due to insufficient training set data in machine learning. The method of the present invention can achieve energy supply and demand balance and low-cost power transmission, realize timely energy storage during power trough periods, and meet the normal charging requests of electric taxis.

[0082] The present invention fully considers the issue of privacy protection and does not require local data of electric taxis and charging stations to complete energy prediction. In addition, a multi-attention mechanism weighted method is introduced when aggregating the model, which greatly improves the prediction accuracy.

[0083] Example:

[0084] See Figure 2 , the urban electric taxi energy prediction system specifically includes: the first-layer federated learning system 11 includes a communication network module 11, a local charging station module (GRU) 12 and an electric taxi module 13; the second-layer federated learning system 22 includes a charging station provider module 21, an Internet module 22 and a charging station module 23.

[0085] The local training end of the first-layer federated learning is the individual electric taxi module (BYD) 13, and the model aggregation end is the charging station module 23. The electric taxi module (BYD) 13 locally trains its own daily charging behavior model, and sends the relevant parameters generated during the training process to the charging station module 23 through the communication network module 11. The second-layer federated learning system 2 includes a charging station module 23, an Internet module 22, and a charging station provider module 21. The local training end of the second-layer federated learning is the charging station module 23, and the aggregation end is the charging station provider module 21. After obtaining the relevant parameters of the electric taxi module (BYD) 13, the charging station module 23 performs aggregation. The charging station module 23 optimizes its own local model and sends the relevant information to the charging station module 23 supplier through the Internet module 22. The charging station module 23 supplier performs the final global aggregation. Figure 2 As shown, the charging station provider module 21 here can be a server or a national grid.

[0086] In this embodiment, the implementation method of the above-mentioned federated learning composition and aggregation is as follows.

[0087] Construct a two-layer federated learning of electric taxi module (BYD) 13-charging station module 23 and charging station module 23-charging station provider module 21, where the charging station provider module 21 sends the initialization model to the electric taxi module (BYD) 13, the electric taxi module (BYD) 13 trains its own daily charging behavior model, and the generated parameters are sent to the charging station module 23. The charging station module 23 optimizes and aggregates its own local model and sends the gradient information to the charging station provider module 21, and the charging station provider module 21 performs the final global aggregation.

[0088] Define the weight vector of the global model of the charging station provider module 21 as ω csp ; Weight vector of charging station module 23 Weight vector of electric taxi module (BYD) 13 i and j represent the number of electric taxi modules (BYD) 13 and the number of charging station modules 23. Assume that the input matrix is ​​X i ={x1,x2,…,x p}, x p For the charging attribute set of an electric taxi module (BYD) 13 in a certain period, a local network model (Gate Recurrent Unit, GRU) can be established:

[0089] Here, the matrix The size is h×p, b i is the network bias, h is the number of network layers, f i (·) represents the RELU activation function. The loss function of each electric vehicle is defined as:

[0090]

[0091] The batch gradient descent method is used to obtain the weight update method of the electric taxi module (BYD) 13:

[0092]

[0093] Electric taxi module (BYD) charging at a charging station module 23 Gradient information of local network model (GRU) training 13 It will be transmitted to the charging station module 23. In the first layer of federated learning, the charging station module 23 acts as a model aggregator, and the electric taxi module (BYD) 13 outputs the energy demand model for a certain period. Because the charging station module 23 has different attention for each electric taxi module (BYD) 13, we will select different electric taxi module (BYD) 13 parameters for aggregation for different charging station modules 23, which will further solve the problem of unbalanced model quality.

[0094] Select the charging probability statistics and Gaussian distribution probability values ​​as attention to form the aggregation weight. Define E j The probability value of selecting the charging station module 23 for a certain electric vehicle and obtaining the maximum probability of selecting the jth charging station module 23 is:

[0095]

[0096]

[0097] In the formula Represents the number of charging times. The probability of the electric taxi module (BYD) 13 selecting the charging station module 23 in the region also obeys the two-dimensional Gaussian distribution. The parameters generated by the two-dimensional Gaussian distribution model should be introduced in the attention mechanism of statistical probability:

[0098]

[0099] Where: A is a constant, represents the probability of selecting a charging station module 23 in a certain area, L(k) represents the activity point of the electric taxi module (BYD) 13 near a charging station module 23 before charging, and L(j) is the position of a charging station module 23. In summary, the loss function and weight update method of the charging station module 23 are obtained:

[0100]

[0101]

[0102] This can solve the problem of quality imbalance when the charging station module 23 aggregates the local model. In the second-layer federated learning, the charging station provider module 21 aims to minimize the loss function of the local model (charging station module 23). In view of the geographical characteristics of the data distribution of the charging station module 23, the charging station module 23 service provider focuses more on the service area of ​​the charging station module 23 and the correlation between the charging station module 23. When performing the second-layer federated learning aggregation, the geographical fence information, namely the Manhattan distance function, is introduced:

[0103]

[0104] here Represents the central position of a certain area, calculates the Manhattan ratio relationship of all charging station modules 23 and normalizes them, and then we get:

[0105]

[0106] Where: Another aspect is that there are service competition, speed competition and other interaction relationships between the charging station modules 23, and the correlation coefficient between a charging station module 23 and its closest charging station module 23 is defined.

[0107]

[0108] In the formula It represents the combination of the frequency of charging vehicles and the waiting time within a certain period of the charging station module 23. At this time, the loss function of our charging station provider module 21 can be expressed as:

[0109]

[0110] Under the weighted method of the multi-attention mechanism, the model of weight update of the charging station provider module 21 can be obtained:

[0111]

[0112] In the two-layer federated learning, the charging station provider module 21 absorbs the gradient information of the charging station module 23 until the model parameters of the charging station provider module 21 are aggregated, and the global loss function L g (ω csp ) takes the minimum value, and the termination condition is that the prediction error converges to a certain value s or reaches a certain number of iterations T.

[0113] The above description is only a preferred implementation example of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of implementation involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the concept of the application. For example, the above features are replaced with the technical features disclosed in the present application but not limited to those with similar functions.

Claims

1. A method for predicting energy of urban electric taxis, characterized in that: The following steps are involved: S1, training the electric taxi according to the initialization model requirements of the charging station service provider, obtaining a daily charging behavior model of the electric taxi, and transmitting the generated daily charging behavior model parameters to the charging station; S3, the charging station can aggregate the daily charging behavior model parameters of the electric taxis, optimize the local model of the charging station according to the aggregated parameters, transmit the optimized local model parameters to the charging station service provider for full aggregation, obtain the full aggregation prediction model, and use the full aggregation prediction model to predict the energy of the electric taxis; Let the input matrix be X i ={x1,x2,…,x p }, x p For the charging attribute set of the pth electric taxi, a local network model can be established: The batch gradient descent method can be used to obtain the weight update method of the electric taxi: Gradient information for training the local network model of an electric taxi charging at a charging station will be delivered to the charging station; the weight vector of the electric taxi i and j represent the number of electric taxis and charging stations, respectively; Charging probability statistics and Gaussian distribution probability values ​​are used as attention to form aggregation weights; The loss function and weight update method of the charging station: Select charging probability statistics and Gaussian distribution probability values ​​as attention to form aggregation weights; define E j The probability value of selecting a charging station module for an electric vehicle and obtaining the maximum probability of selecting the jth charging station module is: In the formula Represents the number of charging times. The probability of the electric taxi module selecting the charging station module in the area also obeys the two-dimensional Gaussian distribution. The parameters generated by the two-dimensional Gaussian distribution model should be introduced in the attention mechanism of statistical probability: Where: A is a constant, L(k) represents the activity point of the electric taxi module near a charging station module before charging, and L(j) is the location of a charging station module; represents the probability of selecting the jth charging station in a certain area, k is the number of charging stations, and the weight vector of the charging station When performing the second-layer federated learning aggregation, the geo-fence information, namely the Manhattan distance function, is introduced: here Represents the central position of a certain area, calculates the Manhattan ratio of all charging station modules and normalizes them to get: Where: Define the correlation coefficient between a charging station module and its closest charging station module: In the formula represents the combination of the frequency of charging vehicles and waiting time within a certain period of the charging station module; at this time, the loss function of the charging station provider module can be expressed as: Under the weighted method of the multi-attention mechanism, the model of the weight update of the charging station provider module can be obtained:

2. A method for predicting energy of urban electric taxis according to claim 1, characterized in that: The electric taxi module adopts the local model of electric taxi, the communication network module is used for data communication transmission, and the charging station module is used to simulate charging at the charging station.

3. The method for predicting energy of urban electric taxis according to claim 1, characterized in that: The electric taxi model makes local predictions based on the relevant data of the electric taxi’s own charging, and uploads the weight change values ​​to the charging station model through the communication network.

4. A method for predicting energy of urban electric taxis according to claim 1, characterized in that: The local model corresponding to the electric taxi module simulates an operating taxi that is powered solely by electricity.

5. The method for predicting energy of urban electric taxis according to claim 1, characterized in that: The parameters of the daily charging behavior model are obtained through the charging data collection of electric taxi owners.

6. A method for predicting energy of urban electric taxis according to claim 1, characterized in that: The electric taxi module locally trains the daily charging behavior model using a machine learning training method.

7. An urban electric taxi energy prediction system, characterized in that: It includes the first-layer federated learning system and the second-layer federated learning system; The first-level federated learning system includes an electric taxi module, a communication network module, and a local charging station module. The electric taxi module is trained locally based on the local charging station module to obtain a daily charging behavior model. The second-layer federated learning system includes a charging station module, an Internet module, and a charging station provider module. The charging station module is used to obtain the daily charging behavior model parameters of the electric taxi module and then aggregate them. It optimizes the local model of the charging station based on the aggregated parameters and sends the optimized local model parameters to the charging station provider module through the Internet module. The charging station provider module performs the final global aggregation to obtain a fully aggregated prediction model. Let the input matrix be X i ={x1,x2,…,x p }, x p For the charging attribute set of the pth electric taxi, a local network model can be established: The batch gradient descent method can be used to obtain the weight update method of the electric taxi: Gradient information for training the local network model of an electric taxi charging at a charging station will be delivered to the charging station; the weight vector of the electric taxi i and j represent the number of electric taxis and charging stations, respectively; Charging probability statistics and Gaussian distribution probability values ​​are used as attention to form aggregation weights; The loss function and weight update method of the charging station: Select charging probability statistics and Gaussian distribution probability values ​​as attention to form aggregation weights; define E j The probability value of selecting a charging station module for an electric vehicle and obtaining the maximum probability of selecting the jth charging station module is: In the formula Represents the number of charging times. The probability of the electric taxi module selecting the charging station module in the area also obeys the two-dimensional Gaussian distribution. The parameters generated by the two-dimensional Gaussian distribution model should be introduced in the attention mechanism of statistical probability: Where: A is a constant, L(k) represents the activity point of the electric taxi module near a charging station module before charging, and L(j) is the location of a charging station module; represents the probability of selecting the jth charging station in a certain area, k is the number of charging stations, and the weight vector of the charging station When performing the second-layer federated learning aggregation, the geo-fence information, namely the Manhattan distance function, is introduced: here Represents the central position of a certain area, calculates the Manhattan ratio of all charging station modules and normalizes them to get: Where: Define the correlation coefficient between a charging station module and its closest charging station module: In the formula represents the combination of the frequency of charging vehicles and waiting time within a certain period of the charging station module; at this time, the loss function of the charging station provider module can be expressed as: Under the weighted method of the multi-attention mechanism, the model of the weight update of the charging station provider module can be obtained:

Images