Power-transport network charging load prediction method and system, and storage medium

Abstract

The application discloses a power-traffic network charging load prediction method and system, and a storage medium. An electric vehicle basic model, a road network basic model, a charging station basic model, and a power distribution network basic model are constructed. Based on the basic models and historical traffic data, dynamic traffic flow data is generated through a generative adversarial network. A traffic travel demand matrix is constructed based on a start-end point analysis method, and combined with the dynamic traffic flow data and the basic models, a queuing theory algorithm is used to simulate the charging decision, queuing and waiting, and the whole charging service process of the electric vehicle in the charging station. A space-time graph attention network model is constructed, taking the charging demand simulation data, the road network basic model and the charging station basic model as inputs, and through a multi-head graph attention mechanism, the space-time coupling relationship between the traffic network state and the charging demand is obtained, the space-time distribution of the electric vehicle charging load in a future period is predicted, and accurate prediction of the space-time part of the electric vehicle charging load is realized.

Description

Technical Field

[0001] This invention belongs to the field of power grid and transportation network coupling technology. The invention relates to a method, system and storage medium for predicting charging load in a power-transportation network. Background Technology

[0002] With the rapid popularization of electric vehicles, the charging load they introduce into the power system exhibits significant spatiotemporal randomness and uncertainty, posing severe challenges to the planning, safe operation, and optimized scheduling of the distribution network. Electric vehicle charging load is not only affected by complex spatiotemporal factors such as user travel behavior, road networks, and traffic flow conditions, but is also closely related to the service capacity of charging stations.

[0003] Traditional load forecasting methods mostly focus on a single dimension of the power system or only on time series analysis, neglecting the deep coupling and interaction between the transportation network and the power network, as well as the dynamic propagation characteristics of charging load in the spatial dimension, resulting in limited forecast accuracy. In addition, existing methods still have shortcomings in accurately characterizing the randomness of traffic flow and effectively mapping it to the spatiotemporal distribution of charging load, making it difficult to meet the high requirements for forecast accuracy in power-transportation collaborative optimization. Summary of the Invention

[0004] The purpose of this invention is to provide a method, system, and storage medium for predicting charging load in a power-transportation network. It constructs a multi-source basic model framework that integrates electric vehicle characteristics, road network, charging station operating parameters, and distribution network topology. Based on the ST-GAT model, it deeply explores the complex spatiotemporal coupling relationship between the transportation network and charging demand. Finally, through queuing theory simulation and power flow calculation, it achieves accurate prediction of the spatiotemporal distribution of electric vehicle charging load and quantitative assessment of its impact on the carrying capacity of the distribution network.

[0005] The technical solution to achieve the purpose of this invention is as follows:

[0006] A method for predicting charging load in a power-transportation network includes the following steps:

[0007] Construct basic models, which include basic models of electric vehicles, road networks, charging stations, and power distribution networks;

[0008] Based on the basic model and historical traffic data, dynamic traffic flow data is generated through generative adversarial networks.

[0009] Based on the origin-end point analysis method, a traffic demand matrix is ​​constructed. Combined with the dynamic traffic flow data and the basic model, the dynamic traffic flow data is superimposed on the road network basic model to obtain the traffic network. The queuing theory algorithm is used to simulate the entire process of electric vehicles charging decision-making, queuing, and charging service at charging stations.

[0010] A spatiotemporal graph attention network model is constructed. The model takes as input data the entire process of electric vehicle charging decision-making, queuing, and charging service in charging stations, as well as basic road network and charging station models. The model adaptively learns the spatial dependency strength between different nodes in the traffic network through a multi-head graph attention mechanism. Combined with gated cyclic units to capture the long-term pattern of load changes, the spatiotemporal coupling relationship between traffic network state and charging demand is obtained, and the spatiotemporal distribution of electric vehicle charging load in a certain period of the future is predicted.

[0011] In the preferred technical solution, after predicting the spatiotemporal distribution of electric vehicle charging load over a certain period of time in the future, the solution further includes:

[0012] The predicted spatiotemporal distribution data of electric vehicle charging load is input into the basic model of the distribution network in the form of node injected power. By performing power flow calculations, the operating status of the distribution network under various scenarios is solved, the impact of electric vehicle charging load on key operating indicators of the distribution network is quantitatively evaluated, and the carrying capacity of the distribution network is analyzed.

[0013] In the preferred technical solution, constructing the basic model of the electric vehicle includes:

[0014] The simulation covers different types of electric vehicles, including private cars, taxis, and buses, and sets different travel vehicle parameters for each type of vehicle. These vehicle parameters include battery capacity, initial state of charge, energy consumption per unit distance, and rated charging power under fast and slow charging conditions.

[0015] Based on travel chain data, the origin and destination of vehicles, route selection, and the distribution of dwell time at designated points along the route are simulated.

[0016] The state of charge of a vehicle during operation changes with time according to the following formula:

[0017]

[0018] in, and These represent the vehicle's state of charge at times t and t+1, respectively. The battery charge / discharge efficiency coefficient; This refers to the average power consumption per unit distance traveled by the vehicle. This represents the distance traveled by a vehicle within the interval (t, t+1), and its value is set based on dynamic traffic flow data.

[0019] When the vehicle's state of charge is below the preset charging threshold and it is in a rechargeable area, a charging request is triggered.

[0020] In the preferred technical solution, the construction of the basic road network model includes:

[0021] Abstract the road network as an undirected graph. , where U represents the set of network nodes, including the geographical locations of road intersections, key transportation hubs, and charging stations; W represents the set of edges connecting these nodes, representing road segments, with each edge assigned a weight based on the actual distance between nodes, average travel time, or travel cost;

[0022] The adjacency matrix D records the connection relationships and weights between nodes. If node i is directly connected to node j, then... The weight of the edge is 0 or infinity if it is not.

[0023] Dijkstra's algorithm is used to calculate the shortest distance between any two points in the road network, providing a basis for subsequent traffic flow allocation and travel time estimation.

[0024] In the preferred technical solution, the construction of the basic model of the charging station includes:

[0025] In the basic model of the road network, the charging station is modeled as a coupled node between the power distribution network and the road network;

[0026] In addition to the basic geographical location attribute, each charging station node is also associated with the following attributes: service capacity; rated power of each charging pile; and equipment utilization rate. / S, where Let t be the total power actually used by the charging station, S be the sum of the rated power of all charging piles in the charging station; and let S be the numbering information of the distribution network access node connected to it, to establish the mapping relationship between the transportation network and the distribution network.

[0027] Define the key operating parameters of the charging station, including: the number of charging piles, used to determine the number of service channels in the system; the rated charging power of each charging pile; the equipment utilization rate; and the charging service rate, which sets the charging service time for each vehicle to follow a negative exponential distribution, with the mean determined by the vehicle's battery capacity, charging power, and initial state of charge.

[0028] In the preferred technical solution, the construction of the basic model of the power distribution network includes:

[0029] The electrical connection relationships and parameters of the distribution network are established by adopting the IEEE standard power system topology, and the line impedance and transformer parameters are set based on standard values. The network is coupled with the road network through geographical location information.

[0030] In the preferred technical solution, generating dynamic traffic flow data through generative adversarial networks includes:

[0031] A feature tensor composed of multi-source data is constructed. The feature tensor composed of multi-source data includes: a historical traffic flow feature matrix, which includes the traffic flow and speed of each road segment at the same time in the past N days; an adjacency matrix of the road network basic model, which is used to introduce spatial structure information; a time period feature vector, which includes the time of day, weekdays, morning peak, and evening peak; and weather condition data.

[0032] The feature tensor composed of multi-source data is input into the generator of the generative adversarial network; the generator adopts a neural network structure containing graph convolutional layers and fully connected layers to learn a mapping of realistic traffic flow data from the feature tensor composed of random noise and multi-source data.

[0033] The discriminator of the generative adversarial network adopts a graph convolutional network structure to distinguish the authenticity of traffic flow data of individual nodes and the difference between generated traffic flow and real traffic flow in spatial correlation.

[0034] The training process uses the Wasserstein GAN framework, and its optimization objective is to minimize the Wasserstein distance between the generated data distribution and the real data distribution. The loss function is defined as follows:

[0035]

[0036] in, This indicates that the optimization objective of generator G is to minimize the entire expression. The goal of the discriminator is to maximize the entire expression. It is the distribution of real traffic flow data. G is the data distribution output by the generator, and D is the set of 1-Lipschitz functions. The discriminator D must come from a set of functions that satisfy the 1-Lipschitz continuity condition. This indicates that the calculation is performed for all data from the true data distribution. The expected value of sample x from the sampled sample. This indicates that the computation is performed for all data distributions generated. Mid-sample generation Expected value It is the score given by the discriminator D to the real sample x. The discriminator D is used to generate samples The scoring; through adversarial training, the generator is able to produce dynamic traffic flow data that retains the spatiotemporal characteristics of the real road network.

[0037] In the preferred technical solution, a traffic demand matrix is ​​constructed based on the origin-destination analysis method. Combined with the dynamic traffic flow data and the basic model, a queuing theory algorithm is used to simulate the entire process of electric vehicles' charging decisions, queuing, and charging services at charging stations, including:

[0038] Construction of Transportation Demand Matrix: Based on historical traffic data, construct a transportation demand matrix for origin and destination at different time periods. , representing the number of vehicles traveling from the starting point o to the ending point d within the time period t;

[0039] Traffic flow assignment and route selection: The traffic demand matrix is ​​loaded onto the basic road network model. Dynamic road segment travel time is obtained from dynamic traffic flow data. Using a stochastic user equilibrium route selection model, travel demand is assigned to each path in the road network, thereby obtaining the traffic flow, speed and vehicle travel time at different times in the road network.

[0040] Charging behavior triggering and queuing simulation: When the SOC of the simulated vehicle is lower than the threshold during driving, the charging station is selected based on distance, current queuing status, and charging price; the service process within the charging station is simulated using the M / M / C queuing model.

[0041] Vehicle arrival process: The process of a vehicle arriving at each charging station is set to follow a Poisson process with parameter λ, the value of which is dynamically determined by the number of vehicles arriving at that station.

[0042] Service process: The service time of each charging station follows a negative exponential distribution with parameter μ, where μ is related to the charging power and the amount of charging required;

[0043] The system obtains the charging request queue length, average waiting time, and the final start time and required charging amount for each charging station at different time periods.

[0044] In a preferred embodiment, the spatiotemporal graph attention network model includes a spatial attention module and a temporal attention module. The spatial attention module employs a multi-head graph attention mechanism, which performs attention on each node i in the graph at the first... The feature vector of a layer is represented as:

[0045]

[0046] in, This is the concatenation operator for K vectors, where K is the total number of attention heads. For nonlinear activation functions, ∑ j ∈N( i The expression () represents the summation operation on all nodes j in the neighbor set of node i. These are the normalized attention weight coefficients, calculated from the k-th attention head, representing the importance of node j to node i. Let be the trainable weight matrix specific to the k-th attention head. For neighbor node j in the th... l The input feature vector of the layer;

[0047] The time attention module uses gated loop units to capture time dependencies;

[0048] The input to the spatiotemporal graph attention network model is a spatiotemporal sequence tensor with dimensions of time step × number of nodes × features. The features include the predicted number of charging vehicles at each charging station, queue length, average speed of the road segment, and time features. The output of the spatiotemporal graph attention network model is the predicted charging load value of each charging station in the next T time steps.

[0049] Supervised learning uses mean squared error as the loss function, as shown in the following formula:

[0050]

[0051] in, This is the actual load value. These are the model predictions, where N is the number of samples. These are all the trainable parameters of the model, and the loss function is minimized using the Adam optimizer.

[0052] This invention also discloses a power-transportation network charging load forecasting system for implementing the above-mentioned power-transportation network charging load forecasting method, comprising:

[0053] The basic model building module constructs basic models, which include basic models of electric vehicles, basic models of road networks, basic models of charging stations, and basic models of power distribution networks.

[0054] The dynamic traffic flow generation module generates dynamic traffic flow data based on the base model and historical traffic data through a generative adversarial network.

[0055] The charging demand simulation module constructs a traffic travel demand matrix based on the origin-end point analysis method. It combines the dynamic traffic flow data and the basic model, overlays the dynamic traffic flow data onto the road network basic model to obtain the traffic network, and uses queuing theory algorithm to simulate the entire process of electric vehicles charging decision-making, queuing, and charging service at charging stations.

[0056] The charging load prediction module constructs a spatiotemporal graph attention network model. It takes data from the entire process of electric vehicles' charging decisions, queuing, and charging services at charging stations, as well as basic road network and charging station models, as inputs. Through a multi-head graph attention mechanism, it adaptively learns the spatial dependency strength between different nodes in the traffic network and combines it with gated cyclic units to capture long-term patterns of load changes. This yields the spatiotemporal coupling relationship between traffic network status and charging demand, and predicts the spatiotemporal distribution of electric vehicle charging load over a certain period in the future.

[0057] The present invention also discloses a computer storage medium storing a computer program thereon, wherein when the computer executes the computer program, it implements the power-transportation network charging load prediction method described in any of the above claims.

[0058] The present invention also discloses an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor runs the computer program stored in the memory, wherein the computer program, when executed, implements the power-transportation network charging load forecasting method described in any of the above claims.

[0059] Compared with the prior art, the significant advantages of this invention are:

[0060] 1. By constructing a multi-source basic model framework that integrates electric vehicle characteristics, road network, charging station operating parameters, and distribution network topology, dynamic traffic flow data with spatiotemporal randomness that conforms to the real distribution is generated using the WGAN model to overcome data scarcity or noise interference. Based on ST-GAT, the complex spatiotemporal coupling relationship between the traffic network and charging demand is deeply mined. Finally, through queuing theory simulation and power flow calculation, the spatiotemporal distribution of electric vehicle charging load is accurately predicted and its impact on the distribution network carrying capacity is quantitatively assessed. This effectively solves the problems of insufficient characterization of the power-transportation network coupling system and weak uncertainty handling capability of existing prediction methods, and provides reliable technical support for orderly charging of electric vehicles and distribution network carrying capacity assessment.

[0061] 2. This method adaptively learns the spatial dependence strength between different nodes in the traffic network through a multi-head graph attention mechanism, and captures the long-term pattern of load changes by combining a gated cyclic unit, thereby achieving high-precision prediction of the spatiotemporal distribution of charging load in complex coupled systems. The root mean square error of its prediction results is reduced by about 15% on average compared with existing mainstream methods, effectively improving the accuracy and practicality of the prediction model. Attached Figure Description

[0062] Figure 1 This is a flowchart of the power-transportation network charging load forecasting method in this embodiment;

[0063] Figure 2This is a schematic diagram of the road network and distribution network topology coupling model in this embodiment;

[0064] Figure 3 This is a structural diagram of the WGAN model used to generate dynamic traffic flow in this embodiment;

[0065] Figure 4 This is a comparison chart showing the performance of the WGAN of this invention with other traffic flow simulation methods in this embodiment;

[0066] Figure 5 This is a structural diagram of the ST-GAT model used for charging load prediction in this embodiment;

[0067] Figure 6 This is a comparison of the average MSE of the ST-GAT-based prediction method with other prediction models in this embodiment;

[0068] Figure 7 This document provides details of the voltage over-limit node in the power flow calculation in a specific case of this embodiment.

[0069] Figure 8 This is a schematic diagram of the power-transportation network charging load prediction system in this embodiment. Detailed Implementation

[0070] The principle of this invention is as follows: By constructing a multi-source basic model framework that integrates electric vehicle characteristics, road network, charging station operating parameters, and distribution network topology, dynamic traffic flow data with spatiotemporal randomness that conforms to the real distribution is generated using WGAN to overcome data scarcity or noise interference. Based on ST-GAT, the complex spatiotemporal coupling relationship between the traffic network and charging demand is deeply mined. Finally, through queuing theory simulation and power flow calculation, the spatiotemporal distribution of electric vehicle charging load is accurately predicted and its impact on the distribution network carrying capacity is quantitatively assessed. This effectively solves the problems of insufficient characterization of the power-transportation network coupling system and weak uncertainty handling capability of existing prediction methods, providing reliable technical support for orderly charging of electric vehicles and distribution network carrying capacity assessment.

[0071] Example:

[0072] like Figure 1 As shown, a method for predicting charging load in a power-transportation network includes the following steps:

[0073] Construct basic models, which include basic models of electric vehicles, road networks, charging stations, and power distribution networks;

[0074] Based on the basic model and historical traffic data, dynamic traffic flow data is generated through generative adversarial networks.

[0075] Based on the origin-end point analysis method, a traffic demand matrix is ​​constructed. Combined with the dynamic traffic flow data and the basic model, the dynamic traffic flow data is superimposed on the road network basic model to obtain the traffic network. The queuing theory algorithm is used to simulate the entire process of electric vehicles charging decision-making, queuing, and charging service at charging stations.

[0076] A spatiotemporal graph attention network model is constructed. The model takes as input data the entire process of electric vehicle charging decision-making, queuing, and charging service in charging stations, as well as basic road network and charging station models. The model adaptively learns the spatial dependency strength between different nodes in the traffic network through a multi-head graph attention mechanism. Combined with gated cyclic units to capture the long-term pattern of load changes, the spatiotemporal coupling relationship between traffic network state and charging demand is obtained, and the spatiotemporal distribution of electric vehicle charging load in a certain period of the future is predicted.

[0077] In a preferred embodiment, after predicting the spatiotemporal distribution of electric vehicle charging load over a certain future period, the method further includes:

[0078] The predicted spatiotemporal distribution data of electric vehicle charging load is input into the basic model of the distribution network in the form of node injected power. By performing power flow calculations, the operating status of the distribution network under various scenarios is solved, the impact of electric vehicle charging load on key operating indicators of the distribution network is quantitatively evaluated, and the carrying capacity of the distribution network is analyzed.

[0079] Key performance indicators include:

[0080] Voltage deviation: Calculates the deviation between the voltage at each node in the distribution network and the rated voltage. When the node voltage deviation exceeds the rated voltage... or This is considered exceeding the limit;

[0081] Line overload: Calculate the load rate of each branch (actual current / allowable current carrying capacity). When the line load rate continuously exceeds 100%, it is considered an overload.

[0082] System peak-valley difference rate: The ratio of the difference between the daily maximum load and the minimum load to the maximum load is used to assess load fluctuation.

[0083] By analyzing the changes of the above indicators before and after the connection of predicted charging load, the impact of electric vehicle charging load on the carrying capacity of the distribution network is quantitatively assessed.

[0084] In a preferred embodiment, constructing a basic model of an electric vehicle includes:

[0085] The simulation covers different types of electric vehicles, including private cars, taxis, and buses, and sets different travel vehicle parameters for each type of vehicle. These vehicle parameters include battery capacity (unit: kWh), initial state of charge (SOC, range: 0-1), energy consumption per unit distance (unit: kWh / km), and rated charging power (kW) under fast charging and slow charging.

[0086] Based on travel chain data, the origin-destination (OD) of vehicles, route selection, and the distribution of dwell time at designated points (such as residential areas, commercial areas, and workplaces) along the route are simulated.

[0087] The state of charge of a vehicle during operation changes with time according to the following formula:

[0088]

[0089] in, and These represent the vehicle's state of charge at times t and t+1, respectively. The battery charge / discharge efficiency coefficient (typically) When driving (Indicates energy loss). The average power consumption per unit distance traveled by the vehicle (unit: kWh / km). The distance traveled by a vehicle within the interval (t, t+1) is measured in km, and its value is set based on dynamic traffic flow data.

[0090] When the vehicle's state of charge is below the preset charging threshold and it is in a rechargeable area, a charging request is triggered.

[0091] In a preferred embodiment, constructing a basic road network model includes:

[0092] Abstract the road network as an undirected graph. , where U represents the set of network nodes, including the geographical locations of road intersections, key transportation hubs, and charging stations; W represents the set of edges connecting these nodes, representing road segments, with each edge assigned a weight based on the actual distance between nodes, average travel time, or travel cost;

[0093] Using the adjacency matrix D (whose dimension is ) Record the connection relationships and weights between nodes. If node i is directly connected to node j, then... The weight of the edge is 0 or infinity if it is not.

[0094] Dijkstra's algorithm is used to calculate the shortest distance between any two points in the road network, providing a basis for subsequent traffic flow allocation and travel time estimation.

[0095] In a preferred embodiment, constructing the basic model of a charging station includes:

[0096] In the basic model of the road network, the charging station is modeled as a coupled node between the power distribution network and the road network;

[0097] In addition to basic geographical location attributes, each charging station node is associated with the following attributes: service capacity (such as the number of vehicles that can be served simultaneously, i.e., the number of charging piles); rated power of each charging pile; and equipment utilization rate. / S, where Let t be the total power actually used by the charging station, S be the sum of the rated power of all charging piles in the charging station; and let S be the numbering information of the distribution network access node connected to it, to establish the mapping relationship between the transportation network and the distribution network.

[0098] Define the key operating parameters of the charging station, including: the number of charging piles, used to determine the number of service channels in the system; the rated charging power of each charging pile; the equipment utilization rate; and the charging service rate, which sets the charging service time for each vehicle to follow a negative exponential distribution, with the mean determined by the vehicle's battery capacity, charging power, and initial state of charge.

[0099] In a preferred embodiment, constructing the basic model of the distribution network includes:

[0100] The electrical connection relationships and parameters of the distribution network are established by adopting the IEEE standard power system topology, and the line impedance and transformer parameters are set based on standard values. The network is coupled with the road network through geographical location information.

[0101] In a preferred embodiment, generating dynamic traffic flow data via a generative adversarial network includes:

[0102] A feature tensor composed of multi-source data is constructed, which includes: a historical traffic flow feature matrix, which includes the traffic flow and speed of each road segment at the same time in the past N days; an adjacency matrix of the road network basic model, used to introduce spatial structure information; a time period feature vector, which includes the time of day, weekdays, morning peak, and evening peak; and weather condition data (such as categorical variables or specific meteorological indicators such as sunny days, rainy days, and haze).

[0103] The feature tensor composed of multi-source data is input into the generator of the generative adversarial network; the generator adopts a neural network structure containing graph convolutional layers and fully connected layers to learn a mapping of realistic traffic flow data from the feature tensor composed of random noise and multi-source data.

[0104] The discriminator of the generative adversarial network adopts a graph convolutional network structure to distinguish the authenticity of traffic flow data of individual nodes, as well as the differences between generated traffic flow and real traffic flow in spatial correlation (i.e., the coordinated change pattern of traffic flow between road network nodes, such as congestion propagation).

[0105] The training process uses the Wasserstein GAN framework, and its optimization objective is to minimize the Wasserstein distance between the generated data distribution and the real data distribution. The loss function is defined as follows:

[0106]

[0107] in, This indicates that the optimization objective of generator G is to minimize the entire expression. The goal of the discriminator is to maximize the entire expression. It is the distribution of real traffic flow data. G is the data distribution output by the generator, and D is the set of 1-Lipschitz functions. The discriminator D must come from a set of functions that satisfy the 1-Lipschitz continuity condition. This indicates that the calculation is performed for all data from the true data distribution. The expected value of sample x from the sampled sample. This indicates that the computation is performed for all data distributions generated. Mid-sample generation Expected value It is the score given by the discriminator D to the real sample x. The discriminator D is used to generate samples The scoring; through adversarial training, the generator is able to produce dynamic traffic flow data that retains the spatiotemporal characteristics of the real road network.

[0108] In a preferred embodiment, a traffic demand matrix is ​​constructed based on origin-endpoint analysis. Combining the dynamic traffic flow data and the basic model, the dynamic traffic flow data is overlaid onto the road network basic model to obtain a traffic network. A queuing theory algorithm is then used to simulate the entire process of electric vehicles' charging decisions, queuing, and charging services at charging stations, including:

[0109] Construction of Transportation Demand Matrix: Based on historical traffic data, construct origin-destination (OD) transportation demand matrices for different time periods (in hours). , representing the number of vehicles traveling from the starting point o to the ending point d within the time period t;

[0110] Traffic flow assignment and route selection: The traffic demand matrix is ​​loaded onto the basic road network model. Dynamic road segment travel time is obtained from dynamic traffic flow data. Using a stochastic user equilibrium route selection model, travel demand is assigned to each path in the road network, thereby obtaining the traffic flow, speed and vehicle travel time at different times in the road network.

[0111] Charging behavior triggering and queuing simulation: When the SOC of the simulated vehicle is lower than the threshold during driving, the charging station is selected based on distance, current queuing status, and charging price; the service process within the charging station is simulated using the M / M / C queuing model.

[0112] Vehicle arrival process: The process of a vehicle arriving at each charging station is set to follow a Poisson process with parameter λ, the value of which is dynamically determined by the number of vehicles arriving at that station.

[0113] Service process: The service time (i.e., charging time) of each charging station follows a negative exponential distribution with parameter μ, where μ is related to the charging power and the amount of charging required.

[0114] The system obtains the charging request queue length, average waiting time, and the final start time and required charging amount for each charging station at different time periods.

[0115] In a preferred embodiment, the spatiotemporal graph attention network model includes a spatial attention module and a temporal attention module.

[0116] The spatial attention module employs a multi-head graph attention mechanism, which focuses on the attention of each node i in the graph at the first... The feature vector of a layer is represented as:

[0117]

[0118] in, This is the concatenation operator for K vectors, where K is the total number of attention heads. For nonlinear activation functions, ∑ j ∈N( i The expression () represents the summation operation on all nodes j in the neighbor set (including itself) of node i. These are the normalized attention weight coefficients, calculated from the k-th attention head, representing the importance of node j to node i. Let be the trainable weight matrix specific to the k-th attention head. For neighbor node j in the th... l The input feature vector of the layer;

[0119] The temporal attention module uses gated recurrent units (GRUs) to capture temporal dependencies;

[0120] The input to the spatiotemporal graph attention network model is a spatiotemporal sequence tensor with dimensions of time step × number of nodes × features. The features include the predicted number of charging vehicles at each charging station, queue length, average speed of the road segment, and time features. The output of the spatiotemporal graph attention network model is the predicted charging load value (unit: kW) of each charging station in the next T time steps (e.g., the next 24 hours, with 15-minute intervals, for a total of 96 time points).

[0121] Supervised learning uses mean squared error as the loss function, as shown in the following formula:

[0122]

[0123] in, This is the actual load value. These are the model predictions, where N is the number of samples. These are all the trainable parameters of the model, and the loss function is minimized using the Adam optimizer.

[0124] Another embodiment, such as Figure 8 As shown, a power-transportation network charging load forecasting system is used to implement the above-mentioned power-transportation network charging load forecasting method, comprising:

[0125] The basic model building module constructs basic models, which include basic models of electric vehicles, basic models of road networks, basic models of charging stations, and basic models of power distribution networks.

[0126] The dynamic traffic flow generation module generates dynamic traffic flow data based on the base model and historical traffic data through a generative adversarial network.

[0127] The charging demand simulation module constructs a traffic travel demand matrix based on the origin-end point analysis method, and combines the dynamic traffic flow data and basic model to simulate the entire process of electric vehicles' charging decision-making, queuing, and charging service at charging stations using queuing theory algorithms.

[0128] The charging load prediction module constructs a spatiotemporal graph attention network model. It takes data from the entire process of electric vehicles' charging decisions, queuing, and charging services at charging stations, as well as basic road network and charging station models, as inputs. Through a multi-head graph attention mechanism, it adaptively learns the spatial dependency strength between different nodes in the traffic network and combines it with gated cyclic units to capture long-term patterns of load changes. This yields the spatiotemporal coupling relationship between traffic network status and charging demand, and predicts the spatiotemporal distribution of electric vehicle charging load over a certain period in the future.

[0129] Preferably, the power-transportation network charging load forecasting system may also include a distribution network carrying capacity assessment module, which is used to call power flow calculation methods, receive the predicted charging load, assess its impact on the distribution network (key operating indicators such as distribution network node voltage deviation, line overload and system peak-valley difference rate), and obtain analysis results.

[0130] The modules are connected in sequence to form a closed-loop process from basic data input to final distribution network carrying capacity assessment.

[0131] In practical implementation, a typical area of ​​a city is used as the simulation object, which includes 37 road nodes, of which 15 correspond to charging stations. The IEEE 33-node distribution network system is used as the power grid topology. The following steps are included:

[0132] S1. Basic Model Construction Steps: Construct basic models to describe the characteristics of electric vehicles, road network topology, charging station operating parameters, and distribution network topology and electrical parameters, respectively, as the input basis and data carrier of the entire prediction framework.

[0133] This step involves building four basic models, specifically including the following steps:

[0134] S1.1 Constructing a basic electric vehicle model: Simulating three types of electric vehicles: private cars (70%), taxis (20%), and buses (10%). The parameters are set as follows:

[0135] Battery capacity: 60kWh for private cars, 80kWh for taxis, and 200kWh for buses;

[0136] Initial SOC: follows a uniform distribution U(0.4, 0.8);

[0137] Energy consumption per unit distance: 0.15-0.25 kWh / km;

[0138] Charging power: 60kW for fast charging, 7kW for slow charging.

[0139] The state of charge of a vehicle changes with time during operation according to the following formula:

[0140]

[0141] Among them, the charge / discharge efficiency coefficient Take 0.98 (charging) and 0.95 (driving). Output from the dynamic traffic flow generation step (S2).

[0142] When the vehicle's SOC is ≤30%, a charging demand is triggered, and the charging station selection strategy is based on the shortest path algorithm.

[0143] S1.2 Constructing the basic model of the road network: Abstracting the road network into an undirected graph. The node set U contains 37 nodes, of which 15 are where charging stations are located. The weight of the edge set W is the road segment travel time (in seconds). Dijkstra's algorithm is used to calculate the shortest path between OD pairs, and the adjacency matrix D is generated using OpenStreetMap data.

[0144] In this implementation case, the road network of a typical area of ​​a city is obtained and modeled. (See...) Figure 2 The blue part.

[0145] S1.3, Constructing the Basic Model of the Charging Station: Define the key operating parameters of the charging station, mainly including:

[0146] Each charging station is equipped with 4-6 fast charging piles and 6-15 slow charging piles; equipment utilization rate is calculated as follows: / S, where Let t be the total power actually used by the charging station at time t, and S be the sum of the rated power of all charging piles in the charging station.

[0147] The charging service time follows a negative exponential distribution, with the mean μ determined by the vehicle battery capacity, charging power, and initial state of charge.

[0148] In addition, the geographical location attribute of the charging station needs to be configured, and the numbering information of the road access node connected to it and the numbering information of the power distribution network access node connected to it need to be configured according to the geographical location.

[0149] S1.4 Constructing the basic model of the distribution network: Using the IEEE standard power system topology, establish the electrical connection relationships and parameters of the distribution network for subsequent load-bearing capacity analysis.

[0150] In this implementation case, the IEEE 33-bus standard power system topology is adopted. Line impedance and transformer parameters are set based on standard values. The system is coupled with the road network through geographical location information. (See...) Figure 2 The black section contains symbols starting with 'b' representing busbars and symbols starting with 'g' representing generators.

[0151] S2. Dynamic traffic flow generation steps: Using Wasserstein Generative Adversarial Network (WGAN), dynamic traffic flow data that reflects spatiotemporal randomness and uncertainty is generated based on the road network model and historical traffic data.

[0152] In this step, Wasserstein generative adversarial network modeling specifically includes the following steps:

[0153] S2.1 Generator Input: Includes historical traffic flow data: average speed matrix of road segments at 15-minute intervals over the past 30 days (dimension 37 nodes × 96 time steps); road network adjacency matrix (dimension 37 nodes × 37 nodes); external features: time period (morning peak, evening peak), weather (sunny, rainy, snowy).

[0154] S2.2 Generator Structure: Contains a 3-layer Graph Convolutional Network (GCN), with each layer outputting dimensions of [64, 128, 256], using the ReLU activation function. It is used to learn a mapping from a feature tensor composed of random noise and multi-source data to realistic traffic flow data.

[0155] S2.3 Discriminator Structure: The WGAN-GP framework with gradient penalty is adopted. The discriminator is a 2-layer GCN with a hidden layer dimension of 128. A gradient penalty term (λ=10) is added to the loss function.

[0156] S2.4 Training process: The Wasserstein GAN (WGAN) framework is used for training, and its optimization objective is to minimize the Wasserstein distance between the generated data distribution and the real data distribution.

[0157] In this embodiment, the structure diagram of the WGAN model used to generate dynamic traffic flow is shown below. Figure 3 To verify the superiority of this method, a comparison chart of the performance of the WGAN method of this invention with other traffic flow simulation methods is shown below. Figure 4 Applying this method, the Wasserstein distance between the generated data and the real data decreased from the initial value of 1.52 to 0.31, and the spatial correlation coefficient (Pearson correlation coefficient) of the generated traffic flow reached 0.96.

[0158] S3. Charging demand simulation steps: Construct a traffic demand matrix based on origin-destination (OD) analysis and combine it with dynamic traffic flow data (to form a traffic network). Use queuing theory algorithms to simulate the entire process of electric vehicles' charging decisions, queuing, and charging services at charging stations.

[0159] In this step, an OD (Original Demand) matrix needs to be constructed based on historical traffic data, and travel demand needs are allocated to various paths on the road network. Charging behavior is triggered by factors such as the electric vehicle's SOC (State of Charge) status, simulating the entire charging process. The specific steps are as follows:

[0160] S3.1 Construction of OD Demand Matrix: Based on historical traffic data, generate the OD demand matrix for the morning peak (7:00-9:00), in which commuter traffic accounts for 65%.

[0161] S3.2 Traffic flow assignment and route selection: Use a stochastic user equilibrium model to assign OD demand to the road network, and calculate travel time by combining the speed data of dynamic traffic flow generated in step S2.

[0162] S3.3 Charging Behavior Triggering and Queue Simulation: When the vehicle's SOC is below the threshold, the nearest charging station is selected based on the actual queuing situation. The charging behavior is simulated using the M / M / C queuing model (C is the number of charging piles).

[0163] Vehicle arrival process: Assume that the process of a vehicle arriving at each charging station follows a Poisson process with parameter λ, the value of which is dynamically determined by the number of vehicles arriving at that station.

[0164] Service process: The service time (i.e., charging time) of each charging station follows a negative exponential distribution with parameter μ, which is related to the charging power and the amount of charging required.

[0165] This model is used to simulate the entire process of an electric vehicle arriving at a charging station, queuing, and receiving charging services. It outputs the charging request queue length, average waiting time, and the time and amount of charging required for each charging station at different times.

[0166] The transportation network consists of a static road network basic model and dynamic traffic flow superimposed on it. The interaction between the two determines the spatiotemporal distribution of electric vehicles in the road network, thus providing a basis for charging load prediction.

[0167] S4. Charging load prediction step: Construct a spatiotemporal graph attention network (ST-GAT) model, taking the output of the charging demand simulation step, road network topology and charging station information as input, and predicting the spatiotemporal distribution of electric vehicle charging load in a certain period of time by capturing the complex spatiotemporal coupling relationship between traffic network status and charging demand.

[0168] This step, ST-GAT model construction and training, specifically includes the following steps:

[0169] S4.1 ST-GAT Model Construction: The spatiotemporal graph attention network model consists of a spatial attention module and a temporal attention module.

[0170] Spatial Attention Module: Employs a multi-head graph attention mechanism, focusing on the attention of each node i in the graph at the first... The feature vector of a layer is represented as:

[0171]

[0172] in, This is the concatenation operator for K vectors, where K is the total number of attention heads. For nonlinear activation functions, ∑ j∈N( i The expression () represents the summation operation on all nodes j in the neighbor set (including itself) of node i. These are the normalized attention weight coefficients, calculated from the k-th attention head, representing the importance of node j to node i. Let be the trainable weight matrix specific to the k-th attention head. Let be the input feature vector of neighbor node j at layer l.

[0173] Time Attention Module: Uses Gated Cyclic Units (GRUs) with a hidden layer dimension of 64 to capture daily periodicity and trends.

[0174] S4.2 Model Input and Output: The model input is a spatiotemporal sequence tensor with dimensions of (96 time steps × 15 charging stations × 5 features), where the features include the number of charging vehicles, queue length, average road speed, time period encoding, and weather encoding. The model output is the load curve (unit: kW) of each charging station for the next 24 hours (96 time points at 15-minute intervals).

[0175] S4.3 Model Training: Supervised learning is performed using mean squared error (MSE) as the loss function, as shown in the following formula:

[0176]

[0177] in, This is the actual load value. These are the model predictions, where N is the number of samples. These are all trainable parameters of the model, and the Adam optimizer is used to minimize this loss function. The optimizer's initial learning rate is set to 0.001, and the number of training epochs is set to 200.

[0178] In this embodiment, the ST-GAT model structure used for charging load prediction is shown below. Figure 5 To verify the superiority of this method, a comparison chart of the performance of the ST-GAT of this invention with other prediction models is shown below. Figure 6 The average MSE of this method is 32.43 kW, which is better than other methods.

[0179] S5. Distribution network carrying capacity assessment steps: The predicted spatiotemporal distribution data of electric vehicle charging load is input into the distribution network topology model in the form of node injected power. By performing power flow calculation, the impact of electric vehicle charging load on key operating indicators of the distribution network is quantitatively assessed, thereby analyzing the carrying capacity of the distribution network.

[0180] In this step, the predicted charging load of each charging station at each future time needs to be injected into the corresponding access node of the distribution network. The forward-backward substitution method or the Newton-Raphson method is used to calculate the power flow and solve the operating state of the distribution network under each scenario.

[0181] Key performance indicators include:

[0182] S5.1 Voltage Deviation: Calculate the deviation between the voltage at each node in the distribution network and the rated voltage. When the node voltage deviation exceeds the rated voltage... or This is considered exceeding the limit;

[0183] S5.2 Line overload: Calculate the load rate (actual current / allowable current carrying capacity) of each branch. When the line load rate continuously exceeds 100%, it is considered an overload.

[0184] S5.3 System Peak-Valley Difference Rate: The ratio of the difference between the daily maximum load and the minimum load to the maximum load is used to assess load fluctuation.

[0185] By analyzing the changes of the above indicators before and after the connection of predicted charging load, the impact of electric vehicle charging load on the carrying capacity of the distribution network is quantitatively assessed.

[0186] In this embodiment, the output node voltage exceedance status is determined based on the calculation results of the charging load carrying capacity of the distribution network. (See...) Figure 7 .

[0187] In another embodiment, a computer storage medium stores a computer program thereon, which, when executed by a computer, implements the power-transportation network charging load forecasting method described in any of the preceding embodiments.

[0188] The specific implementation method is the one described above, and will not be repeated here.

[0189] In another embodiment, an electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor runs the computer program stored in the memory, wherein the computer program, when executed, implements the power-transportation network charging load forecasting method described in any of the preceding embodiments.

[0190] The specific implementation method is the one described above, and will not be repeated here.

[0191] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A power-transport network charging load prediction method, characterized by, The method comprises the following steps: constructing a basic model, the basic model comprising an electric vehicle basic model, a road network basic model, a charging station basic model, and a power distribution network basic model; generating dynamic traffic flow data based on the basic model and historical traffic data through a generative adversarial network; constructing a traffic travel demand matrix based on a start-end point analysis method, and combining the dynamic traffic flow data and the basic model to superimpose the dynamic traffic flow data on the road network basic model to obtain a traffic network, and simulating the charging decision, queuing and waiting, and charging service whole process of the electric vehicle in the charging station by using a queuing theory algorithm; The space-time graph attention network model is constructed, and the data of the charging decision, queuing and waiting and the whole charging service of the simulated electric vehicle in the charging station, the road network basic model and the charging station basic model are taken as inputs. Through the multi-head graph attention mechanism, the spatial dependence strength between different nodes in the traffic network is adaptively learned, and the long-term regularity of the load change is captured by combining the gated recurrent unit, so that the space-time coupling relationship between the traffic network state and the charging demand is obtained, and the space-time distribution of the electric vehicle charging load in a certain period in the future is predicted. The feature vector of the i-th layer is represented as: in, This is the concatenation operator for K vectors, where K is the total number of attention heads. For nonlinear activation functions, ∑ j ∈N( i The expression () represents the summation operation on all nodes j in the neighbor set of node i. These are the normalized attention weight coefficients, calculated from the k-th attention head, representing the importance of node j to node i. Let be the trainable weight matrix specific to the k-th attention head. For neighbor node j in the th... l The input feature vector of the layer; the time attention module adopts a gated recurrent unit to capture time dependence; the input of the spatio-temporal graph attention network model is a spatio-temporal sequence tensor, which has a dimension of time step × node number × feature, wherein the feature comprises predicted charging vehicle number, queuing length, road section average speed, and time feature of each charging station; the output of the spatio-temporal graph attention network model is predicted charging load value of each charging station in the future T time steps; mean square error is used as a loss function for supervised learning, and the formula is as follows: wherein, is the true load value, is the model predicted value, N is the number of samples, is the model all trainable parameters, this loss function is minimized using the Adam optimizer.

2. The power-traffic network charging load prediction method according to claim 1, characterized in that, after predicting the spatio-temporal distribution of the electric vehicle charging load in the future period, the following steps are further included: the predicted spatio-temporal distribution data of the electric vehicle charging load is connected to the power distribution network basic model in the form of node injection power, the operating state of the power distribution network under each scenario is solved through power flow calculation, the influence of the electric vehicle charging load on the key operating indicators of the power distribution network is quantitatively evaluated, and the carrying capacity of the power distribution network is analyzed.

3. The power-traffic network charging load prediction method according to claim 1, characterized in that, constructing the electric vehicle basic model comprises: simulating different types of electric vehicles, including private cars, taxis and buses, and setting different travel vehicle parameters for each type of vehicle; the vehicle parameters include battery capacity, initial state of charge, unit mileage power consumption, and rated charging power under fast charging and slow charging; based on the trip chain data, the start and end points, path selection and stay time distribution of the vehicle at the set points in the path are simulated; wherein the state of charge of the vehicle changes with time during driving and satisfies the following formula: wherein, and SoC(t) and SoC(t+1) represent the state of charge of the vehicle at time instants t and t+1, respectively; is the battery charge and discharge efficiency coefficient; is the average power consumption per unit of distance traveled by the vehicle; is the distance traveled by the vehicle in the interval (t, t+1), whose value is set according to the dynamic traffic flow data; when the state of charge of the vehicle is lower than a preset charging threshold and is in a chargeable area, the charging demand is triggered.

4. The power-traffic network charging load prediction method according to claim 1, characterized by, constructing the road network basic model comprises: Abstraction of road network as an undirected graph where U represents a set of network nodes, including road intersections, key transportation hubs, and geographical locations of charging stations; W represents a set of edges connecting these nodes, representing road segments, and each edge is assigned a weight, which is derived from the actual distance, average travel time, or travel cost between nodes; The connection relationship and weight between nodes are recorded by using an adjacency matrix D. If node i is directly connected with node j, then is the weight of the edge, otherwise 0 or infinity. Dijkstra algorithm is used to calculate the shortest distance between any two points in the road network, providing a basis for subsequent traffic flow allocation and travel time estimation.

5. The power-traffic network charging load prediction method according to claim 1, characterized in that, constructing the charging station basic model comprises: in the road network basic model, the charging station is modeled as a coupling node of the power distribution network and the road network; Each charging station node is associated with the following attributes in addition to the basic geographical location attribute: service capability; rated power of each charging pile; equipment utilization / S, wherein S is the sum of rated powers of all charging piles of the charging station; and the number information of the access node of the power distribution network connected thereto, to establish a mapping relationship between the traffic network and the power distribution network; key operating parameters of the charging station are defined, including: the number of charging piles for determining the service channel number of the system; the rated charging power of each charging pile; the equipment utilization rate; the charging service rate, the charging service time of each vehicle is set to obey a negative exponential distribution, and the mean value is determined by the vehicle battery capacity, charging power and initial state of charge.

6. The power-traffic network charging load prediction method according to claim 5, characterized in that, constructing the power distribution network basic model comprises: an IEEE standard power system topology is used to establish the electrical connection relationship and parameters of the power distribution network, the line impedance and transformer parameters are set based on standard values, and the power distribution network is coupled with the road network through geographical location information.

7. The power-traffic network charging load prediction method according to claim 1, characterized by, The dynamic traffic flow data is generated by a generative adversarial network, and the generation includes: A multi-source data composed feature tensor is constructed, and the multi-source data composed feature tensor includes: a historical traffic flow feature matrix, the historical traffic flow feature matrix including traffic flow and speed of each road section at the same time of the past N days; an adjacency matrix of a road network basic model, used to introduce spatial structure information; a time period feature vector, the time period feature vector including time of day, weekday, morning peak, evening peak, and weather condition data; The multi-source data composed feature tensor is input into a generator of the generative adversarial network; the generator adopts a neural network structure including a graph convolution layer and a fully connected layer, and is used to learn a mapping of realistic traffic flow data from random noise and the multi-source data composed feature tensor; A discriminator of the generative adversarial network adopts a graph convolution network structure, and is used to distinguish the authenticity of single node traffic flow data and the difference between generated traffic flow and real traffic flow in spatial correlation; A training process is trained by using a Wasserstein GAN framework, and an optimization target is to minimize a Wasserstein distance between a generated data distribution and a real data distribution, and a loss function is defined as follows: wherein, denotes that the optimization objective of the generator G is to minimize the whole expression, denotes that the optimization objective of the discriminator is to maximize the whole expression, is the real traffic flow data distribution, is the data distribution output by the generator G, D is a set of 1-Lipschitz functions, requires that the discriminator D must come from a function set satisfying the 1-Lipschitz continuity condition, denotes the calculation of the expected value for all samples x sampled from the real data distribution , denotes the calculation of the expected value for all generated samples sampled from the generated data distribution , is the score of the real sample x by the discriminator D, is the score of the generated sample by the discriminator D; through the adversarial training, the generator can generate dynamic traffic flow data which retains the real road network spatio-temporal features.

8. The power-traffic network charging load prediction method according to claim 1, characterized in that, The queuing theory algorithm is used to simulate the charging decision, queuing and waiting and the whole charging service process of the electric vehicle in the charging station, and the whole charging service process includes: Traffic demand matrix construction: based on historical traffic data, construct the origin-destination traffic demand matrix in different time periods , represents the number of vehicles traveling from origin o to destination d in time period t; Traffic flow distribution and path selection: a traffic trip demand matrix is loaded onto a road network basic model, dynamic road section passing times are obtained from dynamic traffic flow data, and a random user equilibrium path selection model is used to distribute the trip demand to each path of the road network, so that traffic flow, speed and travel time of vehicles at different times in each time period of the road network are obtained; Charging behavior triggering and queuing simulation: when the simulated vehicle is in the process of driving and the SOC is lower than a threshold value, a charging station is selected based on distance, current queuing condition and charging price; a service process in the charging station is simulated by using an M / M / C queuing model; Vehicle arrival process: the process of vehicles arriving at each charging station is subject to a Poisson process with a parameter λ, and the value of λ is dynamically determined by the number of vehicles arriving at the station; Service process: the service time of each charging pile is subject to a negative exponential distribution with a parameter μ, and μ is related to the charging power and the required charging amount; The charging request queue length, average waiting time, and the time of starting charging and the required charging amount of each charging station at different time periods are obtained.

9. A power-transport network charging load prediction system, characterized by, The power-traffic network charging load prediction method is used to implement the power-traffic network charging load prediction method, and includes: A basic model construction module constructs a basic model, and the basic model includes an electric vehicle basic model, a road network basic model, a charging station basic model, and a power distribution network basic model; A dynamic traffic flow generation module generates dynamic traffic flow data by a generative adversarial network based on the basic model and historical traffic data. The charging demand simulation module constructs a traffic travel demand matrix based on the origin-destination analysis method, and combines the dynamic traffic flow data and the basic model to superimpose the dynamic traffic flow data on the road network basic model to obtain a traffic network, and uses a queuing theory algorithm to simulate the charging decision, queuing waiting and charging service whole process of the electric vehicle in the charging station; The charging load prediction module constructs a spatio-temporal graph attention network model, takes the data of the simulated charging decision, queuing waiting and charging service whole process of the electric vehicle in the charging station, the road network basic model and the charging station basic model as inputs, adaptively learns the spatial dependence strength between different nodes in the traffic network through a multi-head graph attention mechanism, and combines a gated recurrent unit to capture the long-term regularity of the load change to obtain the spatio-temporal coupling relationship between the traffic network state and the charging demand, and predict the spatio-temporal distribution of the electric vehicle charging load in a certain period in the future.

10. A computer storage medium having stored thereon a computer program, characterized in that The computer executes the computer program to realize the power-traffic network charging load prediction method of any one of claims 1-8.

11. An electronic device comprising a memory and a processor, characterized in that The memory stores the computer program, and the processor runs the computer program stored on the memory, and the computer program is executed to realize the power-traffic network charging load prediction method of any one of claims 1-8.

Images