Data-driven spatiotemporal electric vehicle charging routing method and device for ultra-multiple target optimization

Abstract

This invention discloses a data-driven multi-objective optimization method and apparatus for spatiotemporal traffic management in electric vehicle charging. The method includes the following steps: S1, constructing a multi-objective optimization model that simultaneously considers five objectives: minimum energy load, minimum charging or battery leasing cost, minimum electricity price subsidy cost, minimum user waiting time, and maximum utilization rate of charging facilities; S2, using a real-time closed-loop feedback system composed of the multi-objective optimization model and a pre-trained machine learning agent model to calculate the optimized spatiotemporal distribution of vehicle charging; S3, integrating the optimized spatiotemporal distribution of vehicle charging into a navigation system, generating electricity price subsidy costs through a linear feedback model, and guiding vehicles to select the optimal charging strategy by combining navigation strategies and electricity price subsidy costs. This invention can simultaneously and efficiently optimize five conflicting objectives in electric vehicle charging scheduling, achieving reasonable allocation of charging resources and effective guidance of user behavior, and is both real-time and efficient.

Description

Technical Field

[0001] This invention belongs to the field of intelligent battery swapping technology for electric vehicles. Specifically, it relates to a data-driven, multi-objective optimization method and device for spatiotemporal flow diversion in electric vehicle charging. Background Technology

[0002] With the increasing demand for energy solutions to address energy shortages and environmental pollution, new energy vehicles (EVs) are being widely adopted. However, the randomness and uncertainty of charging demand during large-scale grid connection of EVs can lead to grid tidal phenomena, resulting in grid instability, difficulties in distribution network reconfiguration, congestion while charging (swapping) vehicles are waiting, and increased costs associated with deploying charging infrastructure. In recent years, numerous scholars have researched various methods to address these issues, but all have certain limitations, as follows:

[0003] (1) Existing technologies mostly focus on localized research on electric vehicle charging issues, rarely considering the needs of multiple parties. For example, some technologies focus on modeling customer responses to electricity prices, but do not fully consider the needs of multiple stakeholders such as battery swapping stations, governments, and charging vehicles.

[0004] (2) Existing technology designs relatively fixed models that cannot be adjusted according to real-time data, making it difficult to adapt to the needs of the current time, city, and scenario. They cannot support all available machine learning algorithms and have poor versatility and portability.

[0005] (3) Existing methods have slow convergence speed, cannot quickly adjust and optimize strategies for real-time data, and cannot cope with highly dynamic scenarios. Summary of the Invention

[0006] The primary objective of this invention is to overcome the shortcomings and deficiencies of existing technologies and provide a data-driven, multi-objective optimization method for the spatiotemporal flow of electric vehicle charging. This invention can simultaneously and efficiently optimize five conflicting objectives in electric vehicle charging scheduling, achieve reasonable allocation of charging resources and effective guidance of user behavior, effectively solve the problems of uneven allocation of electric vehicle charging resources, long user waiting time, and low utilization rate of charging facilities, and has real-time performance and high efficiency.

[0007] The second objective of this invention is to provide a data-driven, multi-objective optimization spatiotemporal diversion device for electric vehicle charging.

[0008] The objective of this invention is achieved through the following technical solution: a data-driven, multi-objective optimization method for spatiotemporal flow diversion in electric vehicle charging, comprising the following steps:

[0009] S1. Construct a multi-objective optimization model that simultaneously considers five objectives: minimum energy load, minimum charging or battery leasing cost, minimum electricity price subsidy cost, minimum user waiting time, and maximum utilization rate of charging facilities.

[0010] S2. Based on the multi-objective optimization model and the pre-trained machine learning agent model, a real-time closed-loop feedback system is formed to calculate the optimized spatiotemporal distribution of vehicle charging.

[0011] The real-time closed-loop feedback system specifically includes: using a pre-trained machine learning proxy model to dynamically adjust the prediction results based on the output results of the multi-objective optimization model, thereby obtaining the predicted spatiotemporal distribution of vehicle charging; based on the adjusted prediction results, using a multi-objective optimization algorithm to perform real-time collaborative optimization of five objectives, and outputting the optimized spatiotemporal distribution of vehicle charging.

[0012] S3. Integrate the optimized spatiotemporal distribution of vehicle charging into the navigation system, and generate the electricity price subsidy cost through a linear feedback model. Combine the navigation strategy and the electricity price subsidy cost to guide the vehicle to select the optimal charging strategy.

[0013] Preferably, in step S1, the objective functions of the multi-objective optimization model are defined as follows:

[0014] (1) Minimize energy load:

[0015]

[0016] Where f1 is the minimum energy load, P grid (t) represents the total load power at time t. T represents the total scheduling time period;

[0017] (2) Minimize charging or battery leasing costs:

[0018]

[0019] Where f2 is the minimum charging or battery leasing cost, i is the vehicle number sequence number, j is the charging facility number sequence number, t is the time, n is the total number of charging vehicles, m is the total number of charging facilities, and C is the minimum charging or battery leasing cost. ij (t) represents the unit cost of charging or battery leasing, Q ij (t) represents the amount of charge;

[0020] (3) Minimize electricity price subsidy costs:

[0021]

[0022] Where f3 is the minimum electricity price subsidy cost, S ij (t) represents the unit cost of electricity price subsidies;

[0023] (4) Minimize user waiting time:

[0024]

[0025] Where f4 is the shortest waiting time, Wi is the waiting time for the i-th vehicle, ai is the arrival time, and Si is the charging start time;

[0026] (5) Maximize the utilization rate of charging facilities:

[0027]

[0028] Where f5 is the maximum utilization rate of the charging facility, Pj(t) is the actual charging power of the j-th charging facility at time t, and P j rated Rated power,

[0029] Preferably, in step S2, the input to the machine learning proxy model is time-related data and spatial-related data, and the output is the prediction results of the time and spatial distribution of electric vehicle charging demand. The construction method includes the following steps:

[0030] Step a: Collect time-related data and spatial-related data and perform data preprocessing;

[0031] Step b, Model selection and construction: Build an LSTM-CNN hybrid model and perform optimization to obtain an optimized machine learning proxy model;

[0032] Step c, Model Training: Train the initial machine learning agent model using historical spatiotemporal distribution data of charging demand; then continuously optimize and update the initial machine learning agent model using real-time spatiotemporal distribution data of charging demand, so that the machine learning agent model can dynamically output the latest prediction results based on the decision variables of the model output by multiple objectives.

[0033] Preferably, in step a, the time-related data includes charging records, external time factors and their corresponding time series matrix A and external time factor vector.

[0034] The spatially related data includes the geographical location information of charging facilities, vehicle driving trajectory data, geospatial data and its corresponding geographical location matrix G, vehicle driving trajectory matrix H and regional geospatial matrix S.

[0035] Preferably, in step a, the data preprocessing specifically includes:

[0036] Convert the time data in the time series matrix A into a unified format, and then compare the converted time series matrix A with the external time factor vector. Extract features to obtain new feature vectors

[0037] The coordinate data of the geographic location matrix G is converted to the same coordinate system, and the vehicle trajectory matrix H and the regional geospatial matrix S are then quantized.

[0038] Preferably, step b specifically includes the following steps:

[0039] The LSTM network is constructed by including an input layer, a hidden layer, and an output layer. Preprocessed temporally relevant data is input into the LSTM network, the hidden state is calculated using the basic unit formula of the LSTM network, and the output of the LSTM network serves as a preliminary prediction result of the temporal distribution.

[0040] The CNN network is constructed by including an input layer, convolutional layers, pooling layers, and an output layer. The preprocessed spatially related data is used to construct a spatial data matrix. Convolutional kernels are used to perform convolution operations to extract spatial features. Pooling operations are performed on the output of the convolutional layers to reduce the data dimensionality. The output of the CNN network serves as the preliminary prediction result of the spatial distribution.

[0041] Model optimization includes replacing LSTM with Gated Recurrent Units (GRUs) and Temporal Separable Convolutions, and replacing CNN with Depth-Separable Convolutions and Channel Attention Mechanisms. A distributed computing strategy is also adopted, where the central node continuously performs computations, while the edge nodes use lightweight offline models for computation and periodically update the model from the central node.

[0042] The fusion component involves fusing the outputs of the LSTM and CNN. The fused result is then processed through a fully connected layer to generate the final prediction results for the temporal and spatial distribution of electric vehicle charging demand.

[0043] Preferably, in step S2, the real-time collaborative optimization of the five objectives using a multi-objective optimization algorithm specifically refers to: using a multi-objective optimization algorithm based on adaptive reference line hybrid projection distance to solve the problem, including the following steps:

[0044] S21. Initialization: Randomly generate the initial population P. t Initialize the reference vector set R, and set the maximum number of evaluations t. max This is a termination condition;

[0045] S22, Evolutionary Cycle: This includes generating offspring populations through crossover and mutation operations, and converting the parent population P... t and offspring population Q t Merge to form a new population; establish a uniformly distributed initial reference line and associate population individuals with the nearest reference line; calculate fitness and select the population individual with the best fitness; adaptively adjust the reference line according to the association relationship;

[0046] S23. Termination and Output: When the maximum number of evaluations t is reached...max This terminates the evolutionary cycle and returns to the final population, thus obtaining the optimized solution set.

[0047] Preferably, step S22 specifically includes:

[0048] (1) Fitness calculation based on hybrid projection distance:

[0049]

[0050] Wherein, φ(p) i ) is the fitness function, the first term Representative solution p i Projected distance on the reference line ri; second term Representative solution p i The global distance to the minimum coordinate point zmin of all solutions, multiplied by a weight. t is the iteration time, t max The maximum number of evaluations is M, where M is the number of targets; the third item Representative solution p i The perpendicular distance to the reference line ri;

[0051] (2) Adaptive reference line adjustment:

[0052] Sparse reference line splitting: If it is related to reference line r i The number of associated solutions is less than the average. A new reference line r is generated around it. new The formula is:

[0053]

[0054] Where r is a random vector and M is the number of targets;

[0055] Dense reference lines merge: If it merges with reference line r i The number of associated solutions is greater than the average. Then select the reference line r with the fewest associated solutions among its adjacent reference lines. j To merge, the formula is:

[0056]

[0057] Preferably, in step S3, the linear feedback model is:

[0058]

[0059] Where s0 is the baseline subsidy amount and s0≥0, N ij (t) represents the actual charging demand, which is predicted using a machine learning proxy model.

[0060] Preferably, step S3 specifically includes:

[0061] The spatial distribution of optimized vehicle charging time and space distribution is integrated into the navigation system to guide vehicles in real time to select the optimal charging route to the nearest and least occupied charging facility; and the electricity price subsidy cost is generated through a linear feedback model to guide vehicles to select the optimal charging time based on the electricity price subsidy cost.

[0062] A data-driven, multi-objective optimization spatiotemporal diversion device for electric vehicle charging, comprising:

[0063] The first building module is used to build a multi-objective optimization model with the goals of minimizing energy load, minimizing charging or battery leasing costs, minimizing electricity price subsidy costs, minimizing user waiting time, and maximizing the utilization rate of charging facilities.

[0064] The second building module is used to construct a real-time closed-loop feedback system based on a multi-objective optimization model and a pre-trained machine learning agent model, and to calculate the optimized spatiotemporal distribution of vehicle charging.

[0065] The real-time closed-loop feedback system specifically includes: using a pre-trained machine learning proxy model to dynamically adjust the prediction results based on the output results of the multi-objective optimization model, thereby obtaining the predicted spatiotemporal distribution of vehicle charging; based on the adjusted prediction results, using a multi-objective optimization algorithm to perform real-time collaborative optimization of five objectives, and outputting the optimized spatiotemporal distribution of vehicle charging.

[0066] The guidance module is used to integrate the optimized spatiotemporal distribution of vehicle charging into the navigation system, and generate the electricity price subsidy cost through a linear feedback model, and guide the vehicle to select the optimal charging strategy based on the electricity price subsidy cost.

[0067] The present invention has the following advantages and effects compared with the prior art:

[0068] (1) This invention provides a data-driven, multi-objective optimization method for the spatiotemporal flow of electric vehicle charging. First, a multi-objective optimization model (five-objective optimization model) is constructed, which simultaneously considers five conflicting objectives: energy load, charging or battery leasing costs, electricity price subsidy costs, user waiting time, and charging facility utilization rate. It has the advantages of being more detailed and comprehensive, and being able to adapt to more complex scenarios (including comprehensively considering the balance and interests of multiple conflicting objectives such as charging facility operators, governments, and users), thus overcoming the shortcomings of existing technologies that lack consideration of multiple needs. By combining the multi-objective optimization model and the machine learning proxy model to form a real-time closed-loop feedback system, it can effectively optimize the five objectives in electric vehicle charging scheduling, realize the rational allocation of charging resources and the effective guidance of user behavior, and effectively solve problems such as uneven distribution of electric vehicle charging resources, long user waiting time, and low utilization rate of charging facilities, providing technical support for the popularization of electric vehicles and the intelligent management of charging networks.

[0069] (2) This invention employs a big data-based machine learning method (LSTM-CNN hybrid model) to construct a machine learning proxy model. Optimizations are made to the LSTM+CNN architecture, including replacing LSTM with gated recurrent units (GRUs) and temporally separable convolutions to reduce the number of LSTM parameters, and replacing CNN with depthwise separable convolutions and channel attention mechanisms to reduce computational load. This allows the machine learning proxy model to be embedded in multi-objective evolutionary algorithms, providing a foundation for continuous and rapid optimization of these algorithms. Simultaneously, a distributed computing strategy is adopted, where the central node (cloud) performs computation while the edge side uses a distilled, lightweight offline model. This enables continuous optimization at each edge side (charging facility) using only low-computing-power modules. This machine learning proxy model can predict the spatiotemporal distribution of charging demand for all vehicles connected to the system in real time based on the latest input data, exhibiting strong practicality and wide applicability.

[0070] (3) The present invention uses a multi-objective evolutionary optimization algorithm based on adaptive reference line hybrid projection distance, which can converge quickly in complex high-dimensional objective space, thereby realizing the real-time and continuous optimization iteration of the multi-objective optimization model. It can adapt to high dynamic scenarios (providing car owners with the best charging time, charging route and charging facilities 24 hours a day, and providing charging facility operators or the government with dynamic and currently most suitable charging strategies and electricity price subsidy schemes), ensuring real-time performance and efficiency, and overcoming the shortcomings of existing technologies such as slow convergence speed and difficulty in dealing with high dynamic scenarios. Attached Figure Description

[0071] Figure 1 This is a flowchart illustrating the data-driven, multi-objective optimization method for spatiotemporal diversion of electric vehicle charging according to the present invention.

[0072] Figure 2 This is a logical schematic diagram of the data-driven, multi-objective optimization method for spatiotemporal diversion of electric vehicle charging according to the present invention.

[0073] Figure 3 This is a schematic diagram of the spatiotemporal distribution of charging demand based on machine learning and the real-time optimization framework integrating "cloud-edge-device" in this invention. Detailed Implementation

[0074] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0075] Example 1

[0076] like Figure 1 The diagram shows a data-driven, multi-objective optimization method for spatiotemporal flow management in electric vehicle charging, including the following steps:

[0077] S1. Construct a multi-objective optimization model that simultaneously considers five objectives: minimum energy load, minimum charging or battery leasing cost, minimum electricity price subsidy cost, minimum user waiting time, and maximum utilization rate of charging facilities.

[0078] S2. Based on the multi-objective optimization model and the pre-trained machine learning agent model, a real-time closed-loop feedback system is formed to calculate the optimized spatiotemporal distribution of vehicle charging.

[0079] The real-time closed-loop feedback system specifically includes: using a pre-trained machine learning proxy model to dynamically adjust the prediction results based on the output results of the multi-objective optimization model, thereby obtaining the predicted spatiotemporal distribution of vehicle charging; based on the adjusted prediction results, using a multi-objective optimization algorithm to perform real-time collaborative optimization of five objectives, and outputting the optimized spatiotemporal distribution of vehicle charging.

[0080] S3. Integrate the optimized spatiotemporal distribution of vehicle charging into the navigation system, and generate the electricity price subsidy cost through a linear feedback model. Combine the navigation strategy and the electricity price subsidy cost to guide the vehicle to select the optimal charging strategy.

[0081] Specifically, this invention provides a data-driven, multi-objective optimization method for the spatiotemporal flow of electric vehicle charging. First, a multi-objective optimization model (five-objective optimization model) is constructed, which simultaneously considers five conflicting objectives: energy load, charging or battery leasing costs, electricity price subsidy costs, user waiting time, and charging facility utilization rate. This method has the advantages of being more detailed and comprehensive, and able to adapt to more complex scenarios (including comprehensively considering the balance and interests of multiple conflicting objectives such as charging facility operators, governments, and users), thus overcoming the shortcomings of existing technologies that lack consideration of the needs of multiple parties.

[0082] Specifically, this invention integrates the spatial distribution of optimized vehicle charging time and space distribution into the navigation system, guiding vehicles in real time to select the optimal charging route to the nearest and least crowded charging facility or station, thus avoiding congestion caused by a large number of vehicles crowding into the same charging station; and generates electricity price subsidy costs through a linear feedback model, guiding vehicles to select the optimal charging time based on the electricity price subsidy costs, thus avoiding congestion at charging stations and grid overload caused by a large number of vehicles charging at the same time.

[0083] This multi-objective optimization model guides navigation systems to direct drivers to the optimal charging facilities at the most suitable time. It also leverages charging subsidy policies at various charging stations to guide drivers in choosing the best charging facility at the most appropriate time. These two spatiotemporal guidance methods not only maximize the interests of car owners but also minimize the costs for subsidized charging facilities (private) or government (state-owned).

[0084] In addition, such as Figure 2 As shown, this invention combines a multi-objective optimization model and a machine learning proxy model to form a real-time closed-loop feedback system. This system can predict and dynamically adjust spatiotemporal distribution data in real time and continuously, effectively optimizing the five conflicting objectives in electric vehicle charging scheduling. It achieves reasonable allocation of charging resources and effective guidance of user behavior, effectively solving problems such as uneven distribution of electric vehicle charging resources, long user waiting time, and low utilization rate of charging facilities. This provides technical support for the popularization of new energy vehicles and the intelligent management of charging networks.

[0085] In step S1, the objective functions of the multi-objective optimization model are defined as follows:

[0086] (1) Minimize energy load:

[0087]

[0088] Where f1 is the minimum energy load, P grid (t) represents the total load power at time t. T represents the total scheduling time period;

[0089] (2) Minimize charging or battery leasing costs:

[0090]

[0091] Where f2 is the minimum charging or battery leasing cost, i is the vehicle number sequence number, j is the charging facility number sequence number, t is the time, n is the total number of charging vehicles, m is the total number of charging facilities, and C is the minimum charging or battery leasing cost. ij (t) represents the unit cost of charging or battery leasing, Q ij (t) represents the amount of charge;

[0092] (3) Minimize electricity price subsidy costs:

[0093]

[0094] Where f3 is the minimum electricity price subsidy cost, S ij (t) represents the unit cost of electricity price subsidies; S ij (t) is calculated based on the linear feedback model, the expression of which is:

[0095]

[0096] Where s0 is the baseline subsidy amount and s0≥0, N ij (t) represents the actual charging demand, which is predicted using a machine learning proxy model.

[0097] (4) Minimize user waiting time:

[0098]

[0099] Where f4 is the shortest waiting time, W i Let a be the waiting time for the i-th vehicle. i For arrival time, S i Start charging time;

[0100] (5) Maximize the utilization rate of charging facilities:

[0101]

[0102] Where f5 is the maximum utilization rate of charging facilities, P j (t) represents the actual charging power of the j-th charging facility at time t, P j rated Rated power,

[0103] In step S2, the input to the machine learning proxy model is time-related data and spatial-related data, and the output is the prediction results of the time and spatial distribution of electric vehicle charging demand. The construction method includes the following steps:

[0104] Step a: Collect time-related data and spatial-related data and perform data preprocessing;

[0105] Step a specifically includes the following process:

[0106] (1) Modeling data collection:

[0107] Time-related data collection: The time-related data includes charging records and external time factors. A corresponding time series matrix A and an external time factor vector are constructed using the charging records and external time factors, respectively.

[0108] Specifically, in this embodiment, charging records are obtained from the charging operator's database and organized into a time series matrix A∈R. n×2 Where n is the number of charging records, t si t represents the start time of the i-th charge. ei This represents the end time of the i-th charge.

[0109] Seasonal information is obtained from meteorological departments, and weekdays / rest days are determined from calendar data; these external time factors are then integrated into a vector. For example s represents the season (numerical codes 1-4 represent spring, summer, autumn and winter), w represents the workday / restday (0 for restday, 1 for workday), and h represents the time of day (0-23 represent the number of hours).

[0110] Spatial data collection: The spatial data includes the geographical location information of charging facilities, vehicle driving trajectory data, and geospatial data. The corresponding geographical location matrix G, vehicle driving trajectory matrix H, and regional geospatial matrix S are constructed based on the geographical location information, vehicle driving trajectory data, and geospatial data, respectively.

[0111] Specifically, in this embodiment, the charging location and route data of the vehicle owner are obtained from the electronic map provider; the geographical location information of the charging facilities is obtained from the charging facility operation platform to construct the geographical location matrix G, i.e., the coordinate matrix G∈R. k×2 Where k is the number of charging facilities, x i and y i These are the longitude and latitude coordinates of the i-th charging facility, respectively.

[0112] Vehicle trajectory data is acquired through in-vehicle equipment or relevant traffic data platforms and organized into a vehicle trajectory matrix, i.e., a three-dimensional matrix H∈R. l×2×p Where l is the number of vehicle trajectory points, 2 represents the coordinate dimensions (longitude and latitude), p is the number of vehicles, and H... i,j,k This represents the j-th coordinate of the i-th travel trajectory point of the k-th vehicle (j=1 represents longitude, j=2 represents latitude).

[0113] Obtain geospatial data such as population density and land use type of the region from the government's geographic information department, and construct a regional geospatial matrix S∈R. q×r Where q is the number of regions, and r is the number of geospatial data types, such as d i1 d represents the population density of the i-th region. i2This indicates the land use type of the i-th region (numerical codes 1-3 represent commercial area, residential area, and industrial area).

[0114] (2) Data preprocessing

[0115] Time-related data preprocessing: Converting the time data in the time series matrix A into a unified timestamp format or a suitable relative time scale; performing feature engineering on the time data, for example from... Extract more features from the transformed time series matrix A, such as calculating the duration t of each charging session. ei -t si This, combined with external time factors, forms a new feature vector to better represent time patterns.

[0116] Spatial data preprocessing: Necessary coordinate transformations are performed on the coordinate data in the geographic location matrix G (if the coordinate systems are inconsistent) to ensure all coordinates are in the same coordinate system; the regional geospatial matrix S and the vehicle trajectory matrix H are quantized. For example, the geographic region is divided into grids, and information such as the number of charging facilities (according to G) and the number of vehicle trajectory points (according to H) within each grid is collected to reconstruct a spatial data representation more suitable for the model input.

[0117] Step b, Model selection and construction: Build an LSTM-CNN hybrid model and optimize it to obtain an optimized machine learning proxy model;

[0118] Step b specifically includes the following processes:

[0119] (1) Constructing the LSTM network, including the input layer, hidden layer, and output layer:

[0120] The input layer is constructed by combining preprocessed time-related data (including features extracted from charging records and external time factors) into an input vector. Input into the LSTM network;

[0121] An LSTM network consists of multiple LSTM units, and the computation of its hidden layers is based on the fundamental unit formula of LSTM. For example, the forget gate... Input gate Cell state update Output gate h t =o t ⊙tanh(C t ), where W f b f W i b i W C b CW o b o These are the corresponding weight matrix and bias vector, respectively, where σ is the sigmoid function. h t-1 It is the hidden state from the previous moment. It is the input time feature vector at time t;

[0122] After calculations at multiple time steps, the LSTM output y LSTM It could be the hidden state h at the last moment. T (Assuming the time series length is T) or the result after passing through a fully connected layer, this output will serve as a preliminary prediction of the time distribution.

[0123] (2) Constructing the CNN network, including the input layer, convolutional layer, pooling layer, and output layer:

[0124] The input layer is constructed by taking the preprocessed spatial data (such as the density of charging facilities and population density after gridding) and building it into a spatial data matrix X∈R. H×W×D Where H and W represent the height and width of the spatial data matrix (e.g., the number of rows and columns of a gridded geographic region), respectively, and D represents the number of channels (e.g., different spatial features);

[0125] The CNN part includes convolutional layers, pooling layers, etc. Convolutional layer calculation: Let the convolutional kernel K∈R... k×k×D×M Where k is the size of the convolution kernel, and M is the number of convolution kernels (output channels), the convolution operation... Where Y∈R H′×W′×M It is the output of the convolutional layer, H′=H-k+1, W′=W-k+1, b m This is the bias of the m-th output channel. After operations such as pooling layers (e.g., max pooling), the final CNN output y is obtained. CNN This output will serve as a preliminary prediction of the spatial distribution.

[0126] (3) LSTM-CNN model optimization:

[0127] The LSTM-CNN model may be computationally expensive or slow, so LSTM is replaced with Gated Recurrent Units (GRUs) and Temporal Separable Convolutions to reduce the number of parameters; CNN is replaced with Depth-Separable Convolutions and Channel Attention Mechanisms to reduce computational cost.

[0128] Simultaneously, a distributed computing strategy is adopted, namely a time-division strategy where the central node continuously performs computations while the edge nodes use lightweight offline models. The edge side (such as the computing units of each charging facility) uses a lightweight offline model, and the distilled (or pruned) lightweight offline model is periodically updated from the central node to ensure the model's up-to-dateness and consistency. Based on this distributed computing strategy, an integrated "cloud-edge-device" real-time optimization framework is constructed, such as... Figure 3 As shown in the diagram, the cloud is responsible for all model calculations and iterations, periodically distributing the distilled offline models to the corresponding units (charging facilities). The edge devices are responsible for periodically uploading the latest data and feedback to the central cloud model, and periodically downloading the offline models to enable low-cost, rapid inference and optimization at the terminal. The terminal is responsible for periodically uploading and updating spatiotemporal data, and downloading the optimal charging strategy to the navigation system or charging app, etc. This data-driven, real-time optimization "cloud-edge-device" layered deployment framework enables each edge device to continuously optimize using only low-computing-power modules.

[0129] (4) Constructing the fusion component:

[0130] The output y of LSTM LSTM and the output y of CNN CNN To perform fusion, for example using a simple splicing method, the fused output y = [y LSTM ,y CNN This fusion result will be further processed through a fully connected layer and other operations to obtain the final prediction result Y of the temporal and spatial distribution of electric vehicle charging demand.

[0131] Step c, Model Training: Train the initial machine learning agent model using historical spatiotemporal distribution data of charging demand; then continuously optimize and update the initial machine learning agent model using real-time spatiotemporal distribution data of charging demand, so that the machine learning agent model can dynamically output the latest prediction results based on the decision variables output by the model based on multiple objectives (such as the electricity price subsidy cost of all charging stations at time t, and the recommended charging station location for each vehicle at time t).

[0132] Specifically, in this embodiment, the core large model is trained and deployed in the cloud, enabling global model iteration at the minute level. At the edge, the model is locally updated based on regional charging demand and traffic conditions to achieve low energy consumption and second-level model update speeds. At the terminal (vehicle, charging station), the latest offline model is updated on demand in real time via high-speed 5G network to provide terminal results. This integrated "cloud-edge-terminal" deployment allows the model and method to adapt to real-time changes in vehicle location and charging demand in highly dynamic scenarios with lower energy consumption and computational requirements.

[0133] Specifically, in this embodiment, the optimized model architecture includes a temporal feature extraction part (GRU plus temporally separable convolution reduces the computational cost of temporal feature extraction), a spatial feature extraction part (depth-separable convolution plus channel attention mechanism reduces the computational complexity of spatial feature extraction), a feature fusion part, and a distributed deployment strategy (distributed computing strategy combined with model distillation ensures real-time inference capability of edge nodes). Therefore, the above optimization method can significantly reduce the computational cost and resource consumption of the LSTM-CNN model, while improving the model's real-time performance and practicality.

[0134] Furthermore, after model training, the final model, namely the pre-trained machine learning agent model, can predict the spatiotemporal distribution of charging demand for all vehicles connected to the system based on the latest time-related and spatial-related data. This demand is real-time and spatiotemporally distributed, thus it can adapt to cities of different sizes, special scenarios such as the Spring Festival travel rush, and seasonal changes in traffic and charging demand, and has broad application prospects.

[0135] In step S2, the real-time collaborative optimization of the five objectives using a multi-objective optimization algorithm specifically refers to: using a multi-objective evolutionary optimization algorithm based on adaptive reference line hybrid projection distance for optimization, including the following steps:

[0136] S21. Initialization: Randomly generate the initial population P. t (Population size is S), initialize the reference vector set R = {r1, r2, ..., r...} S Set the maximum number of evaluations t. max This is a termination condition;

[0137] S22, Evolutionary Cycle:

[0138] 1) Generate offspring populations through crossover and mutation operations, and then use the parent population P... t and offspring population Q t Merging to form new populations;

[0139] 2) Establish uniformly distributed initial reference lines and associate population individuals with the nearest reference line:

[0140] For each individual in the population, calculate its distance to each reference quantity, associate each individual with the nearest reference line, and record the association relationship;

[0141] 3) Fitness calculation and selection:

[0142] That is, for each reference line, the mixed projective distance fitness of its associated population individuals is calculated. The fitness based on the mixed projective distance is calculated as follows:

[0143]

[0144] Wherein, φ(p) i ) is the fitness function, the first term Representative solution p i At reference line r i Projected distance on; second item Representative solution p i To the smallest coordinate point z of all solutions min The global distance, multiplied by a weight t is the iteration time, t max The maximum number of evaluations is M, where M is the number of targets; the third item Representative solution p i To reference line r i The vertical distance;

[0145] Select the most fit individuals from the population to join the next generation.

[0146] 4) Adaptive reference line adjustment: This involves adaptively adjusting the reference line based on the current population's relationship with the reference line.

[0147] Sparse reference line splitting: If it is related to reference line r i The number of associated solutions is less than the average. A new reference line r is generated around it. new The formula is:

[0148]

[0149] Where r is a random vector and M is the number of targets;

[0150] Dense reference lines merge: If it merges with reference line r i The number of associated solutions is greater than the average. Then select the reference line r with the fewest associated solutions among its adjacent reference lines. j To merge, the formula is:

[0151]

[0152] S23. Termination and Output: When the maximum number of evaluations t is reached... max This terminates the evolutionary cycle and returns to the final population, thus obtaining the optimized solution set.

[0153] Specifically, this invention uses a multi-objective evolutionary optimization algorithm based on adaptive reference line hybrid projection distance, which can converge quickly in complex high-dimensional objective spaces. This enables the multi-objective optimization model to perform real-time and continuous optimization iterations, adapting to highly dynamic scenarios (providing car owners with the best charging time, charging routes, and charging facilities 24 hours a day, and providing charging facility operators or governments with dynamic and currently most suitable charging strategies and electricity price subsidy schemes). It ensures real-time performance and efficiency, overcoming the shortcomings of existing technologies such as slow convergence speed and difficulty in dealing with highly dynamic scenarios.

[0154] Example 2

[0155] A data-driven, multi-objective optimization spatiotemporal diversion device for electric vehicle charging, comprising:

[0156] The first building module is used to build a multi-objective optimization model that simultaneously takes into account five objectives: minimum energy load, minimum charging or battery leasing cost, minimum electricity price subsidy cost, minimum user waiting time, and maximum utilization rate of charging facilities.

[0157] The second building module is used to form a real-time closed-loop feedback system based on the multi-objective optimization model and the pre-trained machine learning agent model to calculate the optimized spatiotemporal distribution of vehicle charging.

[0158] The real-time closed-loop feedback system specifically includes: using a pre-trained machine learning proxy model to dynamically adjust the prediction results based on the output results of the multi-objective optimization model, thereby obtaining the predicted spatiotemporal distribution of vehicle charging; based on the adjusted prediction results, using a multi-objective optimization algorithm to perform real-time collaborative optimization of five objectives, and outputting the optimized spatiotemporal distribution of vehicle charging.

[0159] The guidance module is used to integrate the optimized spatiotemporal distribution of vehicle charging into the navigation system, and generate the electricity price subsidy cost through a linear feedback model. Combining the navigation strategy and the electricity price subsidy cost, it guides the vehicle to select the optimal charging strategy.

[0160] The above embodiments are preferred embodiments of the present invention and are not intended to limit the present invention. Any changes or other equivalent substitutions made without departing from the technical solution of the present invention are included within the protection scope of the present invention.

Claims

1. A data-driven, multi-objective optimization method for spatiotemporal flow diversion in electric vehicle charging, characterized in that: Includes the following steps: S1. Construct a multi-objective optimization model that simultaneously considers five objectives: minimum energy load, minimum charging or battery leasing cost, minimum electricity price subsidy cost, minimum user waiting time, and maximum utilization rate of charging facilities. There are conflicting interests among the five objectives, and it is necessary to solve for a Pareto optimal solution set that can balance the objectives. S2. Based on the multi-objective optimization model and the pre-trained machine learning agent model, a real-time closed-loop feedback system is formed to calculate the optimized spatiotemporal distribution of vehicle charging. The real-time closed-loop feedback system specifically includes: using a pre-trained machine learning proxy model to dynamically adjust the prediction results based on the output results of the multi-objective optimization model, thereby obtaining the predicted spatiotemporal distribution of vehicle charging; based on the adjusted prediction results, using the multi-objective optimization algorithm to perform real-time collaborative optimization of the five objectives, and outputting the optimized spatiotemporal distribution of vehicle charging, i.e., outputting a set of Pareto optimal solutions. S3. Integrate the optimized vehicle charging time and space distribution into the navigation system, and generate the electricity price subsidy cost through a linear feedback model. Combine the navigation strategy and the electricity price subsidy cost to guide the vehicle to select the optimal charging strategy. In step S2, the real-time collaborative optimization of the five objectives using a multi-objective optimization algorithm specifically refers to solving the problem using a multi-objective optimization algorithm based on adaptive reference line hybrid projection distance, including the following steps: S21. Initialization: Randomly generate the initial population P. t Initialize the reference vector set R, and set the maximum number of evaluations t. max This is a termination condition; S22, Evolutionary Cycle: This includes generating offspring populations through crossover and mutation operations, and converting the parent population P... t and offspring population Q t Merge to form a new population; establish a uniformly distributed initial reference line and associate population individuals with the nearest reference line; calculate fitness and select the population individual with the best fitness; adaptively adjust the reference line according to the association relationship; S23. Termination and Output: When the maximum number of evaluations t is reached... max This terminates the evolutionary cycle and returns to the final population, thus obtaining the optimized solution set.

2. The data-driven, multi-objective optimization method for spatiotemporal diversion of electric vehicle charging according to claim 1, characterized in that, In step S1, the objective functions of the multi-objective optimization model are defined as follows: (1) Minimize energy load: Where f1 is the minimum energy load, P grid (t) represents the total load power at time t. T represents the total scheduling time period; (2) Minimize charging or battery leasing costs: Where f2 is the minimum charging or battery leasing cost, i is the vehicle number sequence number, j is the charging facility number sequence number, t is the time, n is the total number of charging vehicles, m is the total number of charging facilities, and C is the minimum charging or battery leasing cost. ij (t) represents the unit cost of charging or battery leasing, Q ij (t) represents the amount of charge; (3) Minimize electricity price subsidy costs: Where f3 is the minimum electricity price subsidy cost, S ij (t) represents the unit cost of electricity price subsidies; (4) Minimize user waiting time: Where f4 is the shortest waiting time, W i Let a be the waiting time for the i-th vehicle. i For arrival time, S i Start charging time; (5) Maximize the utilization rate of charging facilities: Where f5 is the maximum utilization rate of charging facilities, P j (t) represents the actual charging power of the j-th charging facility at time t. Rated power, 3. The data-driven, multi-objective optimization method for spatiotemporal diversion of electric vehicle charging according to claim 1, characterized in that, In step S2, the input to the machine learning proxy model is time-related data and spatial-related data, and the output is the prediction results of the time and spatial distribution of electric vehicle charging demand. The construction method includes the following steps: Step a: Collect time-related data and spatial-related data and perform data preprocessing; Step b, Model selection and construction: Build an LSTM-CNN hybrid model and perform optimization to obtain an optimized machine learning proxy model; Step c, Model Training: Train the initial machine learning agent model using historical spatiotemporal distribution data of charging demand; then continuously optimize and update the initial machine learning agent model using real-time spatiotemporal distribution data of charging demand, so that the machine learning agent model can dynamically output the latest prediction results based on the decision variables of the model output by multiple objectives.

4. The data-driven, multi-objective optimization method for spatiotemporal diversion of electric vehicle charging according to claim 3, characterized in that, In step a, the time-related data includes charging records, external time factors and their corresponding time series matrix A and external time factor vector. The spatially related data includes the geographical location information of charging facilities, vehicle driving trajectory data, geospatial data and its corresponding geographical location matrix G, vehicle driving trajectory matrix H and regional geospatial matrix S.

5. The data-driven, multi-objective optimization method for spatiotemporal diversion of electric vehicle charging according to claim 4, characterized in that, In step a, the data preprocessing specifically includes: Convert the time data in the time series matrix A into a unified format, and then compare the converted time series matrix A with the external time factor vector. Extract features to obtain new feature vectors The coordinate data of the geographic location matrix G is converted to the same coordinate system, and the vehicle trajectory matrix H and the regional geospatial matrix S are then quantized.

6. The data-driven, multi-objective optimization method for spatiotemporal diversion of electric vehicle charging according to claim 4, characterized in that, Step b specifically includes the following steps: The LSTM network is constructed by including an input layer, a hidden layer, and an output layer. Preprocessed temporally relevant data is input into the LSTM network, the hidden state is calculated using the basic unit formula of the LSTM network, and the output of the LSTM network serves as a preliminary prediction result of the temporal distribution. The CNN network is constructed by including an input layer, convolutional layers, pooling layers, and an output layer. The preprocessed spatially related data is used to construct a spatial data matrix. Convolutional kernels are used to perform convolution operations to extract spatial features. Pooling operations are performed on the output of the convolutional layers to reduce the data dimensionality. The output of the CNN network serves as the preliminary prediction result of the spatial distribution. Model optimization includes replacing LSTM with Gated Recurrent Units (GRUs) and Temporal Separable Convolutions, and replacing CNN with Depth-Separable Convolutions and Channel Attention Mechanisms. A distributed computing strategy is also adopted, where the central node continuously performs computations, while the edge nodes use lightweight offline models for computation and periodically update the model from the central node. The fusion component involves fusing the outputs of the LSTM and CNN. The fused result is then processed through a fully connected layer to generate the final prediction results for the temporal and spatial distribution of electric vehicle charging demand.

7. The data-driven, multi-objective optimization method for spatiotemporal diversion of electric vehicle charging according to claim 1, characterized in that, Step S22 specifically includes: (1) Fitness calculation based on hybrid projection distance: Wherein, φ(p) i ) is the fitness function, the first term Representative solution p i At reference line r i Projected distance on; second item Representative solution p i To the smallest coordinate point z of all solutions min The global distance, multiplied by a weight t is the iteration time, t max The maximum number of evaluations is M, where M is the number of targets; the third item Representative solution p i To reference line r i The vertical distance; (2) Adaptive reference line adjustment: Sparse reference line splitting: If it is related to reference line r i The number of associated solutions is less than the average. A new reference line r is generated around it. new The formula is: Where r is a random vector and M is the number of targets; Dense reference lines merge: If it merges with reference line r i The number of associated solutions is greater than the average. Then select the reference line r with the fewest associated solutions among its adjacent reference lines. j To merge, the formula is:

8. The data-driven, multi-objective optimization method for spatiotemporal diversion of electric vehicle charging according to claim 2, characterized in that, In step S3, the linear feedback model is: Where s0 is the baseline subsidy amount and s0≥0, N ij (t) represents the actual charging demand, which is predicted using a machine learning proxy model.

9. The data-driven, multi-objective optimization method for spatiotemporal diversion of electric vehicle charging according to claim 1, characterized in that, Step S3 specifically includes: The spatial distribution of optimized vehicle charging time and space distribution is integrated into the navigation system to guide vehicles in real time to select the optimal charging route to the nearest and least occupied charging facility; and the electricity price subsidy cost is generated through a linear feedback model to guide vehicles to select the optimal charging time based on the electricity price subsidy cost.

10. A data-driven, multi-objective optimization spatiotemporal diversion device for electric vehicle charging, applied to the method of claim 1, characterized in that, include: The first building module is used to build a multi-objective optimization model with the goals of minimizing energy load, minimizing charging or battery leasing costs, minimizing electricity price subsidy costs, minimizing user waiting time, and maximizing the utilization rate of charging facilities. The second building module is used to construct a real-time closed-loop feedback system based on a multi-objective optimization model and a pre-trained machine learning agent model, and to calculate the optimized spatiotemporal distribution of vehicle charging. The real-time closed-loop feedback system specifically includes: using a pre-trained machine learning proxy model to dynamically adjust the prediction results based on the output results of the multi-objective optimization model, thereby obtaining the predicted spatiotemporal distribution of vehicle charging; based on the adjusted prediction results, using a multi-objective optimization algorithm to perform real-time collaborative optimization of five objectives, and outputting the optimized spatiotemporal distribution of vehicle charging. The guidance module is used to integrate the optimized spatiotemporal distribution of vehicle charging into the navigation system, and generate the electricity price subsidy cost through a linear feedback model, and guide the vehicle to select the optimal charging strategy based on the electricity price subsidy cost.

Images