New energy automobile charging scheduling method and system based on deep learning

Abstract

The invention discloses a new energy automobile charging scheduling method and system based on deep learning, and the method comprises the steps: carrying out the preprocessing of data, and obtaining multi-source data; a self-attention mechanism of Transform is combined with a GNN graph neural network to establish a deep learning model, a time sequence is processed, a topological relation between geographic distribution of charging piles and a power grid load is modeled through GNN, and spatio-temporal joint features are output; meta-learning is introduced to dynamically adjust the weight of the spatio-temporal joint feature according to a real-time environment, and a charging demand prediction index is obtained; and solving an optimal charging pile distribution scheme through MIP mixed integer programming according to the charging demand prediction index, and generating a charging scheduling scheme by adopting a greedy algorithm. The total charging cost is reduced, the average waiting time of users is shortened, and the power grid load variance is reduced.

Description

Technical Field

[0001] This invention relates to the field of new energy vehicle charging scheduling technology, and in particular to a new energy vehicle charging scheduling method and system based on deep learning. Background Technology

[0002] With the rapid increase in the penetration rate of new energy vehicles, the contradiction between charging demand and grid capacity and charging pile layout is becoming increasingly prominent. Existing charging scheduling methods rely heavily on single charging records or location information at the data level, ignoring key influencing factors such as grid load and battery health status, resulting in incomplete data support. At the model level, they often use single time series or spatial modeling methods, making it difficult to take into account the temporal periodicity and spatial correlation of charging demand. Moreover, the optimization objectives often focus on a single dimension, which can easily lead to grid load imbalance, excessive operating costs, and low scheduling efficiency. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing a new energy vehicle charging scheduling method and system based on deep learning.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned deep learning-based new energy vehicle charging scheduling method: The system acquires user historical charging records, vehicle battery health status, real-time location information, power grid load data, and meteorological data, and preprocesses the data to obtain multi-source data. A deep learning model is established by combining the self-attention mechanism of Transformer with GNN graph neural network. Transformer is used to process time series, and GNN is used to model the topological relationship between the geographical distribution of charging piles and the power grid load, and output spatiotemporal joint features. Meta-learning is introduced to dynamically adjust the weights of the spatiotemporal joint features based on the real-time environment, thereby obtaining the charging demand prediction index. Based on the charging demand prediction index, the optimal charging pile allocation scheme is solved by MIP mixed integer programming, and a greedy algorithm is used to generate a charging scheduling scheme.

[0005] Furthermore, in the aforementioned deep learning-based new energy vehicle charging scheduling method, the acquisition of user historical charging records, vehicle battery health status, real-time location information, power grid load data, and meteorological data, followed by data preprocessing to obtain multi-source data, including: Acquire user's historical charging records, vehicle battery health status, real-time location information, power grid load data, and meteorological data; Linear interpolation is used to fill time series data, while mode filling is used to fill categorical data. The 3σ criterion is used to identify outliers in grid load and charging volume, and invalid data is removed in combination with business logic. The Min-Max normalization method was used to uniformly map battery temperature, grid load, and location coordinate data to the [0,1] interval; Data is synchronized based on timestamps and uniformly adjusted to a 1-minute time granularity. The real-time location of the vehicle is matched with the coordinates of the charging station to calculate the straight-line distance from the vehicle to each charging pile and the estimated travel time, thus obtaining multi-source data.

[0006] Furthermore, in the aforementioned deep learning-based new energy vehicle charging scheduling method, the step of combining the self-attention mechanism of Transformer with GNN graph neural network to establish a deep learning model, using Transformer to process time series data, and GNN to model the topological relationship between the geographical distribution of charging piles and the power grid load, outputting spatiotemporal joint features, including: The standardized time series data is split according to the variable dimension, and each variable is mapped into a 64-dimensional vector through the embedding layer of Transformer, and position encoding information is added; A three-layer self-attention module was constructed, with eight attention heads set in each layer. By calculating the attention weights between data at different time steps, the pattern of charging demand changing over time was captured, and the correlation between multiple variables was explored. The self-attention output passes through a two-layer fully connected Feed Forward network, and through residual connections and layer normalization operations, outputs a temporal feature vector.

[0007] Furthermore, in the aforementioned deep learning-based new energy vehicle charging scheduling method, the step of combining the self-attention mechanism of Transformer with GNN graph neural network to establish a deep learning model, using Transformer to process time series data, and GNN to model the topological relationship between the geographical distribution of charging piles and the power grid load, outputting spatiotemporal joint features, including: Construct a heterogeneous graph of vehicle-charging pile-grid nodes, where nodes include vehicle nodes, charging pile nodes, and grid nodes; A two-layer GCN graph convolutional network is used as the core. The first layer aggregates the neighbor information of each node, and the second layer performs deep processing on the aggregated features to capture the indirect relationships between multiple nodes. Global average pooling is performed on the features of each node output by GCN to compress the features of the heterogeneous graph into a spatial feature vector.

[0008] Furthermore, in the aforementioned deep learning-based new energy vehicle charging scheduling method, the step of combining the self-attention mechanism of Transformer with GNN graph neural network to establish a deep learning model, using Transformer to process time series data, and GNN to model the topological relationship between the geographical distribution of charging piles and the power grid load, outputting spatiotemporal joint features, including: The temporal feature vector output by the Transformer is concatenated with the spatial feature vector output by the GNN to obtain the initial fused features; The correlation weights between various dimensions of temporal and spatial features are calculated using a single-head self-attention module, and spatiotemporal joint features are output.

[0009] Furthermore, in the aforementioned deep learning-based new energy vehicle charging scheduling method, the introduction of meta-learning dynamically adjusts the weights of the spatiotemporal joint features according to the real-time environment to obtain a charging demand prediction index, including: Environmental features are extracted from multi-source data, including current weather type, time period type, and real-time power grid load level, to obtain an environmental feature vector; Meta-learning matches historical meta-tasks based on environmental feature vectors, adjusts the weights of each dimension of spatiotemporal joint features based on parameter update rules trained in the inner loop, assigns high weights to features strongly correlated with the current scene, and reduces the weights of weakly correlated features.

[0010] Furthermore, in the aforementioned deep learning-based new energy vehicle charging scheduling method, the optimal charging pile allocation scheme is solved using MIP (Mixed Integer Programming), and a greedy algorithm is employed to generate the charging scheduling scheme, including: The greedy sorting criterion is to combine the charging demand forecast index from high to low and the arrival time at the charging station from early to late. Collect all vehicle information for each charging station, sort the vehicles according to the sorting criteria to form a service queue for the charging stations, calculate the estimated start and end times for each vehicle, and generate a charging scheduling plan.

[0011] Furthermore, in the deep learning-based new energy vehicle charging scheduling system, the new energy vehicle charging scheduling system includes the following modules: The multi-source data acquisition module is used to acquire user historical charging records, vehicle battery health status, real-time location information, power grid load data and meteorological data, and preprocess the data to obtain multi-source data. The joint feature generation module is used to combine the self-attention mechanism of Transformer with GNN graph neural network to build a deep learning model. Transformer is used to process time series, and GNN is used to model the topological relationship between the geographical distribution of charging piles and the power grid load, and output spatiotemporal joint features. The charging demand prediction module is used to introduce meta-learning to dynamically adjust the weights of the spatiotemporal joint features according to the real-time environment, so as to obtain the charging demand prediction index. The scheduling scheme generation module is used to solve the optimal charging pile allocation scheme by MIP mixed integer programming based on the charging demand prediction index, and to generate a charging scheduling scheme by using a greedy algorithm.

[0012] Furthermore, in the deep learning-based new energy vehicle charging scheduling system, the charging demand prediction module includes the following sub-modules: The extraction submodule is used to extract environmental features from multi-source data, including current weather type, time period type and real-time power grid load level, to obtain an environmental feature vector. The adjustment submodule is used for meta-learning to match historical meta-tasks based on environmental feature vectors, and to adjust the weights of each dimension of spatiotemporal joint features based on the parameter update rules of the inner loop training. Features that are strongly related to the current scene are given high weights, while features that are weakly related are given lower weights.

[0013] Furthermore, in the deep learning-based new energy vehicle charging scheduling system, the charging demand prediction module includes the following sub-modules: The sorting submodule is used to combine the charging demand forecast index from high to low and the arrival time at the charging station from early to late to form a greedy sorting criterion. The generation submodule is used to collect all vehicle information for each charging station, sort the vehicles according to sorting criteria to form a service queue for the charging stations, calculate the estimated start and end times of charging for each vehicle, and generate a charging scheduling plan.

[0014] Its beneficial effects lie in the fact that by integrating core data and implementing refined preprocessing, data noise is eliminated, providing multi-dimensional data support for scheduling decisions. Compared with single data source scheduling accuracy improvement, the Transformer and GNN fusion model achieves deep coupling of spatiotemporal features, capturing both temporal patterns such as weekday peak hours and spatial correlations such as charging pile-grid topology, resulting in improved information entropy of spatiotemporal joint features compared to single models. The meta-learning dynamic weighting mechanism can adapt to real-time scenario changes such as weather and load within 50ms, reducing prediction errors in complex scenarios such as high temperatures and peak hours. The combination of multi-objective MIP and greedy algorithm achieves coordinated optimization of cost, waiting time, and load balance under the premise of meeting constraints such as charging pile capacity and grid load, resulting in reduced total charging costs, shortened average user waiting time, and reduced grid load variance. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of the deep learning-based new energy vehicle charging scheduling method in this invention. Figure 2 This is a schematic diagram of the second embodiment of the deep learning-based new energy vehicle charging scheduling method in this invention. Figure 3 This is a schematic diagram of the first embodiment of the deep learning-based new energy vehicle charging scheduling system in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, a deep learning-based new energy vehicle charging scheduling method includes the following steps: Step 101: Obtain user's historical charging records, vehicle battery health status, real-time location information, power grid load data, and meteorological data; preprocess the data to obtain multi-source data. Specifically, in this embodiment, user historical charging records, vehicle battery health status, real-time location information, power grid load data, and meteorological data are acquired. Time series data are filled using linear interpolation, and categorical data are filled using mode filling. Outliers in power grid load and charging amount are identified using the 3σ criterion, and invalid data is removed in conjunction with business logic. Battery temperature, power grid load, and location coordinate data are uniformly mapped to the [0,1] interval using the Min-Max standardization method. Data is synchronized with time based on timestamps and uniformly adjusted to a 1-minute time granularity. The real-time location of the vehicle is matched with the coordinates of the charging station, and the straight-line distance from the vehicle to each charging pile and the estimated travel time are calculated to obtain multi-source data.

[0020] Data is the core foundation for deep learning model training and scheduling decisions. To ensure the accuracy of subsequent charging demand prediction and scheduling plans, a multi-channel, high-timeliness, targeted data collection system needs to be built to address the characteristics of different data, achieving comprehensive coverage and quality control of data dimensions. For users' historical charging records, which serve as key evidence for analyzing user charging habits and behavioral patterns, structured extraction from the core database of the charging operation platform is necessary. In addition to basic information such as user ID, charging start and end time, charging amount, charging pile number, and payment amount, auxiliary information such as electricity price type, peak / valley / flat period, charging mode, and fast / slow charging should also be collected simultaneously. The time span should be selected from approximately 6 months to 1 year to cover the cyclical changes between weekdays and weekends, and peak and off-peak seasons, while avoiding data timeliness degradation due to excessive time. This provides data support for identifying patterns such as concentrated charging after the morning peak and a surge in charging demand on weekend afternoons. The health status of a vehicle's battery directly impacts charging efficiency and safety. Real-time data collection is achieved through the onboard T-BOX terminal, a standard feature in new energy vehicles. This terminal, connected to the vehicle's CAN bus system, accurately acquires core parameters from the Battery Management System (BMS), including State of Charge (SOC), remaining charge, State of Health (SOH), health status, charge / discharge cycle count, individual cell voltage balance, and battery pack temperature distribution. Setting the collection frequency to once every 5 minutes allows for real-time monitoring of battery status changes while avoiding increased terminal power consumption and data redundancy caused by excessively high collection frequencies. The accuracy of the vehicle's real-time location information directly affects the rationality of charging station allocation. Employing a dual-mode GPS and BeiDou positioning module allows for complementary advantages, maintaining positioning accuracy within 10 meters even in complex environments such as urban areas with tall buildings obstructing the view. Combined with mobile 4G / 5G networks, real-time data transmission is achieved. A dynamic collection frequency strategy is employed: collecting data once every 30 seconds while driving accurately tracks the vehicle's trajectory, switching to once every 5 minutes when stationary, effectively reducing network transmission pressure and terminal power consumption. Power grid load data is crucial to preventing grid overload caused by charging scheduling. It needs to be obtained from a dedicated data interface of the power grid dispatch center. In addition to dynamic parameters such as the outgoing load of each substation in the region and the real-time power, voltage, and current of the charging station access points, historical load curves from the past three months also need to be retrieved simultaneously. A sampling frequency of one minute can accurately capture load fluctuation peaks, providing data support for subsequent load balancing scheduling. Meteorological data has a significant impact on both charging demand and battery performance. It is obtained through public APIs from meteorological departments or third-party refined meteorological services. The sampling range is a 1-kilometer radius area centered on the charging station, collecting parameters such as temperature, precipitation, wind force, and sunshine duration in real time. A sampling frequency of once per hour meets the timeliness requirements. Simultaneously, meteorological data from the same period over the past three years is supplemented to provide a basis for analyzing the correlation between decreased battery charging efficiency in high-temperature weather and delayed charging demand in rainy weather.

[0021] The collected multi-source data suffers from issues such as missing values, anomalies, and inconsistent dimensions. Directly inputting these into the model would severely impact training effectiveness and decision accuracy. Therefore, a comprehensive preprocessing strategy involving cleaning, standardization, and alignment is necessary to construct a high-quality standardized dataset. Data cleaning, as the primary preprocessing step, requires the development of refined processing rules for missing and outlier values. For time-series data such as grid load and SOC, which exhibit continuous variation, linear interpolation is used to fill in missing values. This method accurately calculates the missing values ​​by fitting the linear equations of two adjacent valid data points, avoiding trend distortion caused by traditional mean imputation. For categorical data such as charging pile numbers and user IDs, which have discrete values ​​and high repetition rates, mode imputation is used to preserve the data's distribution characteristics to the greatest extent. Outlier handling employs a dual mechanism of statistical testing and business logic verification. First, the 3σ criterion is used to identify outliers in continuous data, removing data exceeding the mean ± 3 standard deviations, such as sudden spikes or drops in grid load or charging volumes far exceeding the battery's rated capacity. Then, a second verification is performed using business logic. The core objective of data standardization is to eliminate the interference of dimensional differences on model training. Parameters from different data sources have significantly different magnitudes. For example, battery temperature is measured in °C (-20~60), grid load is measured in kW (0~1000), and location coordinates are measured in latitude and longitude, with geographically limited ranges. Directly inputting these into the model can lead to excessive bias towards parameters with large magnitudes. The Min-Max standardization method is used, mapping all data to the [0,1] interval by calculating the original data minus the minimum value and the maximum value minus the minimum value. This preserves the relative magnitudes of the data while ensuring equal weight for each parameter during model training, guaranteeing fairness in subsequent feature extraction. Data alignment and integration are crucial for achieving multi-source data association. Global time synchronization is performed based on the timestamp of grid load data. Since grid data is collected once per minute and has the highest stability, user charging records, battery status, vehicle location, and other data are all uniformly adjusted to a 1-minute time granularity. Data collected less than once per minute is padded using forward padding, while data collected more than once per minute is compressed using mean aggregation. Location information processing employs Geographic Information System (GIS) technology, importing the vehicle's real-time latitude and longitude and the charging pile coordinates into the GIS platform. The straight-line distance from the vehicle to each charging pile is calculated using the Haversine formula, and the travel time is corrected by combining real-time traffic data and obtaining data from the map service API. Finally, a three-dimensional multi-source data matrix of time, location, and attributes is formed, with timestamp as the horizontal axis and vehicle-charging pile-power grid-meteorological attributes as the vertical axis, providing structured input for the subsequent spatiotemporal feature extraction of the model.

[0022] Step 102: Combine the self-attention mechanism of Transformer with GNN graph neural network to establish a deep learning model. Use Transformer to process time series and GNN to model the topological relationship between the geographical distribution of charging piles and the power grid load, and output spatiotemporal joint features. Specifically, in this embodiment, the standardized time-series data is split according to variable dimensions. Each variable is mapped to a 64-dimensional vector through the embedding layer of a Transformer, and positional encoding information is added. A three-layer self-attention module is constructed, with eight attention heads in each layer. By calculating the attention weights between data at different time steps, the pattern of charging demand changing over time is captured, and the correlation between multiple variables is explored. The self-attention output passes through a two-layer fully connected Feed Forward network, and through residual connections and layer normalization operations, a time feature vector is output. A heterogeneous graph of vehicle-charging pile-grid nodes is constructed, where nodes include vehicle nodes, charging pile nodes, and grid nodes. A two-layer GCN graph convolutional network is used as the core. The first layer aggregates the neighbor information of each node, and the second layer performs deep processing on the aggregated features to capture indirect correlations between multiple nodes. Global average pooling is performed on the node features output by the GCN to compress the features of the heterogeneous graph into a spatial feature vector. The temporal feature vector output by the Transformer is concatenated with the spatial feature vector output by the GNN to obtain the initial fused features; the correlation weights between the temporal and spatial features in each dimension are calculated through a single-head self-attention module to output the spatiotemporal joint features.

[0023] The overall model architecture is designed to construct a dual-branch parallel fusion architecture, which processes information in the time and space dimensions respectively, and ultimately achieves deep fusion of spatiotemporal features. The model input layer receives preprocessed multi-source data, which is divided into a time series data subset, historical charging records, grid load, battery status time series data, and a spatial correlation data subset, vehicle location, charging pile location, and grid topology data.

[0024] The system uses a dual-branch processing layer. The Transformer branch processes time-series data, while the GNN branch processes spatially correlated data. The outputs of both branches are then processed through a fully connected layer to unify their dimensions before being merged by concatenation.

[0025] The feature output layer activates the fused features using the ReLU activation function, outputting a 256-dimensional spatiotemporal joint feature vector.

[0026] The Transformer time series processing branch focuses on mining time-dimensional dependencies and extracting dynamic features of multivariate time series. Input embedding and location encoding split the standardized time series data by variable dimensions, such as charging amount, SOC, and grid load. The embedding layer maps each variable into a 64-dimensional vector, and then adds location encoding information to identify the time sequence of the data, thus solving the problem that the Transformer model is not sensitive to the time sequence.

[0027] The multi-layer self-attention mechanism is used to construct a 3-layer self-attention module with 8 attention heads in each layer. By calculating the attention weights between data at different time steps, the mechanism captures the pattern of charging demand changes over time, including the peak charging demand after the morning rush hour and before the evening rush hour on weekdays. At the same time, it explores the correlation between multiple variables, such as the correlation between temperature and battery charging efficiency.

[0028] The Feed Forward network and residual connections, with the self-attention output passing through two fully connected Feed Forward networks and a hidden layer dimension of 256, alleviate the gradient vanishing problem during model training through residual connections and layer normalization operations, and finally output a 128-dimensional time feature vector.

[0029] The GNN spatial correlation processing branch models spatial topological relationships using graph structures and extracts correlation features between geography and power grid load. The graph structure is constructed as follows: a heterogeneous graph is built for vehicles, charging piles, and power grid nodes. The nodes include vehicle nodes with attributes such as location, SOC, and battery SOH; charging pile nodes with attributes such as location, rated power, and current idle state; and power grid nodes with attributes such as region and real-time load. The edge construction rules are as follows: the edge weight between vehicles and charging piles is the reciprocal of the driving distance; the edge weight between charging piles and power grid nodes is the proportion of access power; and the edge weight between power grid nodes is the line transmission capacity.

[0030] The graph convolution computation uses a 2-layer GCN, with the graph convolutional network as the core module. The first layer aggregates the neighbor information of each node. For example, the charging pile node aggregates the charging demand information of surrounding vehicle nodes and the load information of power grid nodes. The second layer performs deep processing on the aggregated features to capture the indirect correlation between multiple nodes and the impact of the load change of charging pile A on the access of adjacent charging pile B to the power grid.

[0031] Feature pooling and output: Global average pooling is performed on the features of each node output by GCN to compress the features of the heterogeneous graph into a 128-dimensional spatial feature vector, which is consistent with the dimension of the temporal feature vector of the Transformer branch.

[0032] Spatiotemporal feature fusion employs a concatenation + attention fusion strategy to achieve efficient fusion of temporal and spatial features. The temporal feature vector output by the Transformer is concatenated with the spatial feature vector output by the GNN to obtain an initial fusion feature of 256 dimensions. Subsequently, a single-head self-attention module is used to calculate the correlation weights between the temporal and spatial features in each dimension, strengthen key correlation features, and assign feature weights to peak periods and high-load areas. Finally, the optimized spatiotemporal joint features are output.

[0033] Step 103: Introduce meta-learning to dynamically adjust the weights of spatiotemporal joint features based on the real-time environment to obtain the charging demand prediction index. Specifically, in this embodiment, environmental features are extracted from multi-source data, including the current weather type, time period type, and real-time power grid load level, to obtain an environmental feature vector. Meta-learning matches historical meta-tasks based on the environmental feature vector, and adjusts the weights of each dimension of the spatiotemporal joint features based on the parameter update rules of the inner loop training. Features that are strongly correlated with the current scene are given high weights, while features that are weakly correlated are given lower weights.

[0034] The meta-learning module is built with the goal of quickly adapting to changes in the real-time environment. It consists of a weight adjustment module based on MAML and model-independent meta-learning. The meta-task design treats charging scheduling tasks under different scenarios as meta-tasks. The scenario division is based on weather conditions (rainy / sunny, high / low temperature), time period type (weekday morning peak, weekday evening peak, weekend), and grid load level (low load / medium load / high load). Each meta-task contains historical spatiotemporal characteristic data of that scenario and corresponding actual charging demand tags.

[0035] The meta-training process employs a dual-loop training mechanism: an inner loop and an outer loop. The inner loop focuses on a single meta-task, such as the evening rush hour scenario on a hot workday, adjusting the initial weights of spatiotemporal features through training with a small amount of data. The outer loop updates the parameters of the meta-learning module based on the training results of all meta-tasks, enabling the module to quickly adapt to new scenarios. The training objective is to minimize the demand prediction error under different scenarios.

[0036] Real-time environmental awareness and dynamic weight adjustment: Real-time collection of environmental parameters drives the meta-learning module to complete adaptive weight adjustment. Real-time environmental feature extraction extracts environmental features from preprocessed real-time data, including the current weather type (determined by temperature, precipitation, and other data), time period type (determined by current time and date), and real-time power grid load level (determined by comparing power grid load data with thresholds), forming an environmental feature vector.

[0037] The weights are adjusted quickly. The meta-learning module matches the most similar historical meta-tasks based on the real-time environmental feature vectors and the parameter update rules based on the inner loop training. Within 50ms, the weights of each dimension of the spatiotemporal joint features are adjusted. Features that are strongly correlated with the current scene and battery temperature-related features in high-temperature scenarios are given higher weights, while features that are weakly correlated are given lower weights.

[0038] The charging demand prediction index is calculated by mapping the weighted spatiotemporal joint features to a quantized prediction index. A fully connected layer maps the weighted 256-dimensional feature vector to a 1-dimensional numerical value, and then the Sigmoid function normalizes the value to the [0,10] range to obtain the charging demand prediction index. A higher index indicates a more urgent charging need for the vehicle at the current time and location, with 0-3 representing low demand, 4-7 representing medium demand, and 8-10 representing high demand.

[0039] Step 104: Based on the charging demand prediction index, solve the optimal charging pile allocation scheme using MIP mixed integer programming, and generate a charging scheduling scheme using a greedy algorithm.

[0040] Specifically, in this embodiment, the greedy sorting criterion is formed by superimposing the charging demand prediction index from high to low and the arrival time at the charging station from early to late; all vehicle information of each charging station is collected, and the vehicles are sorted according to the sorting criterion to form a service queue of the charging stations; the estimated start time and end time of charging for each vehicle are calculated to generate a charging scheduling plan.

[0041] Define the constraints and objective function, clarify the core objective and constraint boundaries of the MIP solution, and ensure the feasibility and optimality of the solution. The objective function is to optimize multiple objectives, including minimizing total charging cost, minimizing user waiting time, and balancing grid load. The total charging cost includes peak and off-peak electricity pricing costs and charging pile operating costs. User waiting time is the time from when the vehicle arrives at the charging pile to when charging begins. Grid load balancing is measured by minimizing the variance of grid load in each region. The multiple objectives are transformed into a single objective through weighted summation. The weights are set according to operational needs, such as increasing the load balancing weight during peak periods.

[0042] Constraints include charging pile capacity constraints (a single charging pile can only serve one vehicle at a time, and the charging power cannot exceed the rated power), vehicle range constraints (the energy consumption of the vehicle when driving to the target charging pile cannot exceed the range corresponding to the current SOC), grid load constraints (the total charging power connected to the grid in each region cannot exceed the rated carrying capacity of the grid in that region), and user time constraints (the time window for users to make reservations for charging must be met).

[0043] The optimal charging station allocation is solved using mixed integer programming (MIP). The branch and bound method is used to solve the MIP model to obtain the optimal matching relationship between charging stations and vehicles. Variables are defined as follows: 0-1 variable x_ij, x_ij=1 indicates that the i-th vehicle is allocated to the j-th charging station, x_ij=0 indicates that it is not allocated; continuous variable t_ij, the charging time of the i-th vehicle at the j-th charging station; and integer variable s_ij, the charging start time of the i-th vehicle at the j-th charging station.

[0044] The model is solved by converting the objective function and constraints into the standard MIP form and using a commercial solver, Gurobi, or the open-source solver, PuLP. A solution accuracy threshold is set so that the difference between the optimal solution and the feasible solution is less than 1%. After the solution is completed, the vehicle-charging pile allocation matrix is ​​output, which clearly defines the target charging pile to be allocated to each vehicle.

[0045] A greedy algorithm generates the charging scheduling order. Based on the MIP allocation results, a greedy strategy is used to determine the service order for multiple vehicles at a single charging station. The ranking criteria are set by combining the charging demand forecast index from high to low and the arrival time at the charging station from early to late to form a greedy ranking criterion, giving priority to vehicles with urgent needs and early arrival times, while taking into account both urgency and fairness.

[0046] The scheduling order is generated by collecting information on all vehicles allocated to each charging station and sorting them according to the sorting criteria to form a service queue for that charging station. The estimated start and end times of charging for each vehicle are calculated to ensure that the charging times of vehicles do not overlap and that the power constraints of the charging station are met.

[0047] The scheduling scheme output and feedback integrates the allocation results and scheduling order to form a complete scheduling scheme and provide real-time feedback. It organizes information such as the vehicle-charging pile allocation relationship, the charging start time, end time, and charging power of each vehicle into structured scheduling instructions, which are pushed to the vehicle terminal and charging pile controller through the scheduling platform. At the same time, it monitors the charging pile operation status and vehicle arrival status in real time. If there are abnormalities such as vehicles not arriving on time or charging pile malfunctions, the scheme recalculation process is triggered.

[0048] Its beneficial effects lie in the fact that by integrating core data and implementing refined preprocessing, data noise is eliminated, providing multi-dimensional data support for scheduling decisions. Compared with single data source scheduling accuracy improvement, the Transformer and GNN fusion model achieves deep coupling of spatiotemporal features, capturing both temporal patterns such as weekday peak hours and spatial correlations such as charging pile-grid topology, resulting in improved information entropy of spatiotemporal joint features compared to single models. The meta-learning dynamic weighting mechanism can adapt to real-time scenario changes such as weather and load within 50ms, reducing prediction errors in complex scenarios such as high temperatures and peak hours. The combination of multi-objective MIP and greedy algorithm achieves coordinated optimization of cost, waiting time, and load balance under the premise of meeting constraints such as charging pile capacity and grid load, resulting in reduced total charging costs, shortened average user waiting time, and reduced grid load variance.

[0049] Please see Figure 2 In the deep learning-based new energy vehicle charging scheduling method, the following steps are taken to obtain user historical charging records, vehicle battery health status, real-time location information, power grid load data, and meteorological data, and to preprocess the data to obtain multi-source data: Step 201: Obtain user's historical charging records, vehicle battery health status, real-time location information, power grid load data, and meteorological data; Step 202: Fill time series data with linear interpolation and fill categorical data with mode. Identify outliers in grid load and charging volume using the 3σ criterion and remove invalid data based on business logic. Step 203: Use the Min-Max normalization method to uniformly map the battery temperature, grid load, and location coordinate data to the [0,1] interval; Step 204: Synchronize the data based on the timestamp, uniformly adjust it to a 1-minute time granularity, match the real-time location of the vehicle with the coordinates of the charging station, calculate the straight-line distance from the vehicle to each charging pile and the estimated travel time, and obtain multi-source data.

[0050] The above describes embodiments of the deep learning-based new energy vehicle charging scheduling method of the present invention. Please refer to [link / reference]. Figure 3 In a deep learning-based new energy vehicle charging scheduling system, the system includes the following modules: The multi-source data acquisition module is used to acquire user historical charging records, vehicle battery health status, real-time location information, power grid load data and meteorological data, and preprocess the data to obtain multi-source data. The joint feature generation module is used to combine the self-attention mechanism of Transformer with GNN graph neural network to build a deep learning model. Transformer is used to process time series, and GNN is used to model the topological relationship between the geographical distribution of charging piles and the power grid load, and output spatiotemporal joint features. The charging demand prediction module is used to introduce meta-learning to dynamically adjust the weights of spatiotemporal joint features based on the real-time environment, thereby obtaining the charging demand prediction index. The scheduling scheme generation module is used to solve the optimal charging pile allocation scheme by MIP mixed integer programming based on the charging demand prediction index, and to generate a charging scheduling scheme using a greedy algorithm.

[0051] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based method for scheduling charging for new energy vehicles, characterized in that, The new energy vehicle charging scheduling method includes the following steps: The system acquires user historical charging records, vehicle battery health status, real-time location information, power grid load data, and meteorological data, and preprocesses the data to obtain multi-source data. A deep learning model is established by combining the self-attention mechanism of Transformer with GNN graph neural network. Transformer is used to process time series, and GNN is used to model the topological relationship between the geographical distribution of charging piles and the power grid load, and output spatiotemporal joint features. Meta-learning is introduced to dynamically adjust the weights of the spatiotemporal joint features based on the real-time environment, thereby obtaining the charging demand prediction index. Based on the charging demand prediction index, the optimal charging pile allocation scheme is solved by MIP mixed integer programming, and a greedy algorithm is used to generate a charging scheduling scheme.

2. The deep learning-based new energy vehicle charging scheduling method as described in claim 1, characterized in that, The process involves acquiring user historical charging records, vehicle battery health status, real-time location information, power grid load data, and meteorological data, and preprocessing the data to obtain multi-source data, including: Acquire user's historical charging records, vehicle battery health status, real-time location information, power grid load data, and meteorological data; Linear interpolation is used to fill time series data, while mode filling is used to fill categorical data. The 3σ criterion is used to identify outliers in grid load and charging volume, and invalid data is removed in combination with business logic. The Min-Max normalization method was used to uniformly map battery temperature, grid load, and location coordinate data to the [0,1] interval; Data is synchronized based on timestamps and uniformly adjusted to a 1-minute time granularity. The real-time location of the vehicle is matched with the coordinates of the charging station to calculate the straight-line distance from the vehicle to each charging pile and the estimated travel time, thus obtaining multi-source data.

3. The deep learning-based new energy vehicle charging scheduling method as described in claim 1, characterized in that, The method combines the self-attention mechanism of Transformer with GNN graph neural network to establish a deep learning model. Transformer is used to process time series data, and GNN is used to model the topological relationship between the geographical distribution of charging piles and the power grid load, outputting spatiotemporal joint features, including: The standardized time series data is split according to the variable dimension, and each variable is mapped into a 64-dimensional vector through the embedding layer of Transformer, and position encoding information is added; A three-layer self-attention module was constructed, with eight attention heads set in each layer. By calculating the attention weights between data at different time steps, the pattern of charging demand changing over time was captured, and the correlation between multiple variables was explored. The self-attention output passes through a two-layer fully connected Feed Forward network, and through residual connections and layer normalization operations, outputs a temporal feature vector.

4. The deep learning-based new energy vehicle charging scheduling method as described in claim 1, characterized in that, The method combines the self-attention mechanism of Transformer with GNN graph neural network to establish a deep learning model. Transformer is used to process time series data, and GNN is used to model the topological relationship between the geographical distribution of charging piles and the power grid load, outputting spatiotemporal joint features, including: Construct a heterogeneous graph of vehicle-charging pile-grid nodes, where nodes include vehicle nodes, charging pile nodes, and grid nodes; A two-layer GCN graph convolutional network is used as the core. The first layer aggregates the neighbor information of each node, and the second layer performs deep processing on the aggregated features to capture the indirect relationships between multiple nodes. Global average pooling is performed on the features of each node output by GCN to compress the features of the heterogeneous graph into a spatial feature vector.

5. The deep learning-based new energy vehicle charging scheduling method as described in claim 1, characterized in that, The method combines the self-attention mechanism of Transformer with GNN graph neural network to establish a deep learning model. Transformer is used to process time series data, and GNN is used to model the topological relationship between the geographical distribution of charging piles and the power grid load, outputting spatiotemporal joint features, including: The temporal feature vector output by the Transformer is concatenated with the spatial feature vector output by the GNN to obtain the initial fused features; The correlation weights between various dimensions of temporal and spatial features are calculated using a single-head self-attention module, and spatiotemporal joint features are output.

6. The deep learning-based new energy vehicle charging scheduling method as described in claim 1, characterized in that, The introduced meta-learning dynamically adjusts the weights of the spatiotemporal joint features based on the real-time environment to obtain a charging demand prediction index, including: Environmental features are extracted from multi-source data, including current weather type, time period type, and real-time power grid load level, to obtain an environmental feature vector; Meta-learning matches historical meta-tasks based on environmental feature vectors, adjusts the weights of each dimension of spatiotemporal joint features based on parameter update rules trained in the inner loop, assigns high weights to features strongly correlated with the current scene, and reduces the weights of weakly correlated features.

7. The deep learning-based new energy vehicle charging scheduling method as described in claim 1, characterized in that, The step of solving the optimal charging pile allocation scheme using MIP (Mixed Integer Programming) based on the charging demand prediction index and generating a charging scheduling scheme using a greedy algorithm includes: The greedy sorting criterion is to combine the charging demand forecast index from high to low and the arrival time at the charging station from early to late. Collect all vehicle information for each charging station, sort the vehicles according to the sorting criteria to form a service queue for the charging stations, calculate the estimated start and end times for each vehicle, and generate a charging scheduling plan.

8. A new energy vehicle charging scheduling system based on deep learning, characterized in that, The new energy vehicle charging dispatch system includes the following modules: The multi-source data acquisition module is used to acquire user historical charging records, vehicle battery health status, real-time location information, power grid load data and meteorological data, and preprocess the data to obtain multi-source data. The joint feature generation module is used to combine the self-attention mechanism of Transformer with GNN graph neural network to build a deep learning model. Transformer is used to process time series, and GNN is used to model the topological relationship between the geographical distribution of charging piles and the power grid load, and output spatiotemporal joint features. The charging demand prediction module is used to introduce meta-learning to dynamically adjust the weights of the spatiotemporal joint features according to the real-time environment, so as to obtain the charging demand prediction index. The scheduling scheme generation module is used to solve the optimal charging pile allocation scheme by MIP mixed integer programming based on the charging demand prediction index, and to generate a charging scheduling scheme by using a greedy algorithm.

9. The deep learning-based new energy vehicle charging scheduling system as described in claim 8, characterized in that, The charging demand prediction module Includes the following sub-modules: The extraction submodule is used to extract environmental features from multi-source data, including current weather type, time period type and real-time power grid load level, to obtain an environmental feature vector. The adjustment submodule is used for meta-learning to match historical meta-tasks based on environmental feature vectors, and to adjust the weights of each dimension of spatiotemporal joint features based on the parameter update rules of the inner loop training. Features that are strongly related to the current scene are given high weights, while features that are weakly related are given lower weights.

10. The deep learning-based new energy vehicle charging scheduling system as described in claim 8, characterized in that, The charging demand prediction module includes the following sub-modules: The sorting submodule is used to sort the charging demand forecast index from high to low and the arrival time at the charging station from early to late as a greedy superposition of sorting criteria. The generation submodule is used to collect all vehicle information for each charging station, sort the vehicles according to sorting criteria to form a service queue for the charging stations, calculate the estimated start and end times of charging for each vehicle, and generate a charging scheduling plan.

Images