Charging duration prediction method based on multi-dimensional spatial-temporal characteristics and battery aging physical constraint

Abstract

The invention discloses a charging duration prediction method based on multi-dimensional spatial-temporal characteristics and battery aging physical constraints, and belongs to the field of new energy automobile battery management. The method comprises the following steps: firstly, reading operation discharge data of a vehicle from a vehicle-mounted OBD device, and obtaining a battery capacity label based on a fixed SOC interval; then, constructing a multi-dimensional data sample by utilizing a voltage interval, and converting traditional time sequence data into graph structure data based on correlation among different data; and then, based on the random discharge data and the space-time model, automatically extracting multi-dimensional space-time characteristics representing battery aging, and taking the aging rate of the battery as a physical constraint to accurately predict the battery capacity. And finally, constructing a battery charging duration empirical formula by using a dynamic particle swarm algorithm, and ensuring that the residual charging duration of the battery is accurately predicted under different charging conditions.

Description

Technical Field

[0001] The invention relates to a method for predicting the remaining fast-charging time of a power battery of an in-service new energy vehicle, and belongs to the field of battery management of new energy vehicles. Background Art

[0002] In the daily use of electric vehicles, charging time is one of the core issues that users are concerned about. Especially in the fast charging scenario, accurate prediction of charging time is crucial for users' itinerary planning. Since electric vehicles are prone to "range anxiety" during long-distance driving, that is, users are worried that they will not be able to reach their destination smoothly due to insufficient battery power. Fast charging technology can replenish enough power for the battery in a short time, effectively reduce charging waiting time, and improve travel efficiency and convenience. However, relying solely on fast charging cannot completely solve the problem. Accurate prediction of remaining charging time is also indispensable. Through accurate charging time estimation, users can better arrange charging plans, avoid unnecessary detention, optimize time management, and improve travel experience. In addition, accurate prediction of remaining charging time can also alleviate range anxiety, enhance users' confidence in electric vehicles, and further promote the popularization of fast charging technology. For charging station operators, accurate time estimation helps optimize resource scheduling, reduce congestion of charging piles, and improve overall operational efficiency. Therefore, accurate prediction of remaining charging time is not only a key means to improve user experience, but also an effective solution to the range anxiety problem faced in the promotion of electric vehicles.

[0003] The current battery charging time prediction methods are mainly divided into the following categories. The method based on historical data relies on a large amount of past charging data for modeling and prediction. Although the prediction accuracy can be gradually improved with the increase of data volume, its flexibility is poor. Especially when the charging environment or conditions change, this method is difficult to adapt to new situations, resulting in a significant increase in prediction errors. The method based on physical models can theoretically provide high-precision predictions by modeling the electrochemical characteristics of the battery in detail. However, the complexity of the internal reactions of the battery makes the calculation process of this method very complicated, resulting in high computational costs in real-time applications, and the model parameters are often difficult to accurately obtain in actual operations. The method based on logical algorithms relies on preset rules and logic to predict the charging time. Under stable and controllable conditions, it can usually produce more accurate results, but it is limited to a specific logical framework. Once a situation that does not conform to the preset logic is encountered, the prediction error will increase significantly, reducing the practicality and reliability of the method. In recent years, artificial intelligence technology has developed rapidly and has been effectively applied in the field of battery capacity prediction. The data-driven method can directly establish a battery fast charging remaining time prediction model based on the historical operation data of the battery without exploring its complex decay mechanism. The basic steps of data-driven modeling include extracting characteristic data related to battery capacity decay and selecting algorithms suitable for specific application scenarios. This paper adopts a data-driven approach and proposes a multi-stage hierarchical fusion of multi-dimensional spatiotemporal features and physical constraints of battery aging to predict the fast charging time. This method can accurately predict the remaining time of battery fast charging without the need for extensive battery testing. Summary of the invention

[0004] With the development of the new energy vehicle industry, fast charging technology is gradually becoming an important technical direction. It can improve the use efficiency and charging speed of new energy vehicles, greatly improve the energy replenishment system, and improve the energy replenishment efficiency. In the daily use of electric vehicles, accurately predicting the remaining time of fast charging is crucial to improving the user's charging experience. It can not only relieve the user's waiting anxiety, but also help users arrange their trips reasonably, thereby further optimizing the overall use experience of electric vehicles. However, the fast charging speed or duration is affected by many factors such as charging pile power, battery capacity, and environmental factors, making it difficult to summarize and construct an accurate and universal method to calculate the fast charging time. The present invention proposes a fast charging time prediction method that integrates multi-dimensional spatiotemporal features and physical constraints of battery aging in a multi-stage hierarchical manner. First, the running discharge data of the vehicle is read from the on-board OBD device, and the battery capacity label is obtained based on a fixed SOC interval; then, the multi-dimensional data sample is constructed using the voltage interval, and the traditional time series data is converted into graph structure data based on the correlation between different data; then, based on the random discharge data and spatiotemporal model, the multi-dimensional spatiotemporal features that characterize battery aging are automatically extracted, and the battery aging rate is used as a physical constraint to accurately predict the battery capacity. Finally, a dynamic particle swarm algorithm is used to construct an empirical formula for battery charging time to ensure accurate prediction of the remaining battery charging time under different charging conditions.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] A charging time prediction method based on multi-dimensional spatiotemporal characteristics and physical constraints of battery aging includes the following steps:

[0007] S1: Clean vehicle charging data and obtain capacity tags;

[0008] S11: Obtaining the running and discharging data of the electric vehicle, including: time, total voltage, current, SOC (battery state of charge), single cell maximum voltage, single cell minimum voltage, single cell maximum temperature, single cell minimum temperature, and recording and storing the above running data;

[0009] S12: When cleaning the running data, first use the threshold range to filter out abnormal data; then use the K-nearest neighbor interpolation method to fill in the missing values ​​to ensure the continuity of the time series data; finally use a low-pass filter to smooth the noise data to ensure the stability and reliability of the data;

[0010] S13: When obtaining the capacity tag, first set a fixed SOC interval to avoid the influence of extreme sections on the capacity evaluation and ensure that each calculation covers the same battery working state; then use the current and time data after cleaning to calculate the battery capacity based on Formula 1;

[0011]

[0012] Among them, C a is the calculated battery capacity, Δt is the sampling interval, I is the charging current, t1 and t2 are the start and end time of discharge respectively, SOC t1 is the discharge start SOC data, SOC t2 It is the SOC data for discharge termination;

[0013] S2: Extract multi-dimensional spatiotemporal features;

[0014] S21: construct multidimensional data samples;

[0015] The Q(V) curve, Temp(V) curve, Time(V) curve and IC curve are used as inputs of the data-driven model; the number of acquisition points L is determined according to the starting voltage V_start, the ending voltage V_end and the voltage interval ΔV;

[0016]

[0017] Wherein, ΔV is set to 2 mV, and the boundary value M of each discharge capacity segment is obtained by setting a fixed length n for each discharge segment;

[0018]

[0019] Where c is the step size, and the maximum integer obtained by the function floor(·) does not exceed the input value;

[0020] In practical applications, it is impossible to obtain the upper limit value of each capacity point; to solve this problem, △Q=[Q1,Q2,…,Q n ]–Q1 is used to replace the capacity sequence, and other temperature and time data are processed in the same way; the final multidimensional data sample is expressed as,

[0021]

[0022] Where △Q is the charge change value, V is the voltage value, ΔTemp is the temperature change value, dQ_dV is the IC change value, and n is the length of the collected data;

[0023] S22: construct a multidimensional data graph structure sample according to the multidimensional data sample;

[0024] S23: Extracting spatiotemporal features;

[0025] First, the aggregation function is used in the k convolutional layers of the graph convolutional network to fuse the aggregate information of the neighboring nodes, and then the aggregate information is combined and updated with the current state of the central node. The calculation process is as follows:

[0026]

[0027] in, represents the feature vector of node i in the kth layer, represents the feature vector of node j in the k-1th layer, Indicates that j belongs to the node, that is, all nodes, f a and f c are the aggregation function and update function respectively, N(i) is the node adjacent to i, and σ is the activation function;

[0028] The propagation rule of the kth convolutional layer is:

[0029]

[0030] Among them, Z k is the node layer output vector, ω k is the weight matrix; f k (A,X) represents the graph convolution function, X represents the feature matrix, represents the adjacency matrix, X k The feature matrix of the kth layer, ω k represents the weight matrix of the kth layer;

[0031] The gated recurrent unit is based on Z k As input, the output finally captures the spatiotemporal characteristics of battery aging h t , the calculation formula is:

[0032]

[0033] Among them, g t and r t are the outputs of the update gate and the reset gate respectively; σ g and σ r They are the activation functions of the update gate and the reset gate respectively; and h t Represent the candidate hidden state and the final hidden state respectively; W g , W r and W h are the update gate, reset gate and weight matrix of the hidden layer respectively; b g , b r and They are the bias items of update gate, reset gate and hidden state respectively;

[0034] S3: Establish a capacity prediction model;

[0035] S31: Model input data construction and model initialization;

[0036] When performing capacity prediction based on the GCN-GRU model, GCN stands for graph convolutional network and GRU stands for gated recurrent unit. First, multi-dimensional features are constructed for the input data. The input data is voltage, current, temperature, and time, and graph structure samples are constructed through these features. The constructed graph includes an adjacency matrix and node features. The adjacency matrix represents the correlation between nodes with different features, and the node features represent the specific physical quantity of each node. These graph structure data are passed as input to the graph convolutional network of the model to extract the spatial dependency between features.

[0037] S32: Spatiotemporal feature extraction of GCN-GRU network;

[0038] After completing the construction and initialization of the input data, the data is sent to the GCN layer for spatial feature extraction; the GCN layer will process the graph structure data, and extract the correlation between different feature dimensions by aggregating the information of nodes and neighboring nodes; the convolved graph structure features are passed to the GRU network, and the GRU layer is specifically used to extract dynamic features in the time dimension; GRU captures the capacity change trend of the battery at different time points by processing time series data, and extracts the time series features of the battery aging process; finally, through the fully connected layer, these spatiotemporal features are mapped to the output value of the battery capacity estimation;

[0039] S33: Determine physical constraints on aging rate and model optimization;

[0040] During the model training process, the loss function includes two types: data loss and physical loss. The model is optimized by minimizing the error between the predicted value and the actual battery capacity. The loss function is shown as follows:

[0041]

[0042] Among them, Loss data Represents the data loss function, Loss physics represents the physical feature loss function, Loss represents the total loss, ω1 and ω2 represent the proportional coefficients, and L i represents the number of historical cycles of the i-th data, n is the number of samples, is the capacity value predicted by the model, y i is the true capacity, y0 is the initial capacity value, L i is the number of cycles;

[0043] The Adam optimizer is used in training to accelerate the convergence of the model with an adaptive learning rate, so that it can better fit the changing trend of battery capacity. In order to prevent overfitting, the data set is divided into a training set and a validation set. The model performance is monitored and hyperparameters are adjusted through the validation set to ensure the generalization ability of the model under different charging and discharging conditions, and finally obtain stable and accurate capacity prediction results.

[0044] S4: Predict charging time based on particle filter algorithm;

[0045] S41: Empirical model construction and parameter initialization; In the empirical model of charging capacity prediction, the initial model is first constructed based on the particle filter algorithm, and the parameters are initialized; the model input includes battery capacity, charging SOC range, charging temperature range and charging rate range; the state evolution of these variables is simulated by the particle filter, and the charging process model in multi-dimensional space is constructed;

[0046] The goal of the model is to dynamically update and predict the hidden variables of the charging process through the particle filtering process; the initialization step includes selecting a set of particles to represent the state distribution and estimating the charging time based on the state transition model; the initialization formula of the particle filtering algorithm is:

[0047] p(x0)~N(μ0,∑0)

[0048] Among them, x0 is the initial state variable; μ0 and Σ0 are the mean and variance of the initial state; p(x0) represents the probability distribution of the initial state x0, N(μ0,∑0) represents the normal Gaussian distribution with mean μ0 and covariance matrix ∑0, and ~ represents "obey" or "comply".

[0049] S42: Continuously update the model based on new input data;

[0050] At each time step, the particle filter corrects the weight and state of the particle through the state transfer equation and the observation equation, and finally estimates the actual charging time more accurately. The observation data is continuously input into the model, and the particle filter resamples the particle weight and estimates the actual charging amount in combination with the new observation value. The particle filter formula for state update and observation is:

[0051] p(x t |y 1:t )∝p(y t |x t )∫p(x t |x t-1 )p(x t-1 |y 1:t-1 )dx t-1

[0052] Among them, p(x t y 1:t ) means that all observations y from time 1 to t are obtained 1：t After that, for the state x at time t t The posterior probability distribution of y 1:t represents the set of all observations from time 1 to time t, that is, {y1, y2, ..., y t}, ∝ means proportional to, x t is the state variable at the tth moment, y t is the observed value at the current moment, p(y t |x t ) is the observation model, which represents the state x t Calculate the observation y t The probability of p(x t |x t-1 ) is a state transition model, which indicates the transition probability of the current state based on the previous state.

[0053] Furthermore, the specific method of step S22 is:

[0054] The multi-dimensional time series data is converted into graph structure data. In this process, the time series of each dimension is used as the node feature vector of the graph structure. The edge construction method is based on the correlation between the nodes, and the calculation formula is:

[0055]

[0056] Among them, S pq represents the similarity between the pth and qth nodes; x p and x q are the eigenvectors of the pth and qth nodes respectively; j is the total number of nodes, ||·||2 is the matrix bi-norm;

[0057] The similarities of all nodes are normalized to a uniform scale; the adjacency matrix A of the graph is expressed as:

[0058]

[0059] Among them, A pq represents the element in the p-th row and q-th column of A.

[0060] Furthermore, in step S23, the features are further screened. The screening method is: when the node feature dimension is greater than the set threshold, a portion of elements on the feature vector are randomly shielded to 0, and the discarded ratio is determined by the hyperparameter p; the features after discarding are calculated as follows:

[0061]

[0062] Among them, Z k ' is the feature matrix after the Dropout operation; Dropout(·, p) is the Dropout operation with a discard ratio of p.

[0063] Furthermore, the calculation formula of the optimizer in step S33 is as follows:

[0064]

[0065] Where η is the learning rate, θ t is the current model parameter, m t is the momentum of the gradient, v t is the weighted average of the squared gradients, and ∈ is a small constant to prevent division by zero.

[0066] The beneficial effects of the present invention are:

[0067] The present invention proposes a universal battery pack fast charging remaining time prediction method based on partial discharge data, which can effectively estimate the remaining charging time of electric vehicles, provide important reference information for charging management and driving planning of electric vehicles, and improve charging efficiency and user experience.

[0068] The present invention proposes a method for extracting multidimensional spatiotemporal characteristics of battery health from partial discharge fragments. It can automatically capture the key local characteristics and timing characteristics of the battery degradation process, significantly improving the estimation accuracy of the battery capacity. The model can adapt to battery data with different data distributions and can be effectively used for battery capacity estimation of vehicles with different battery types and battery pack structures.

[0069] This paper proposes a neural network training method that integrates physical information of battery aging. It can simulate the physical phenomena in the battery degradation process and, under the guidance of physical knowledge, use a small amount of data to train a robust model. This method not only considers the dynamic behavior of battery degradation, but also considers the degradation trend, thereby more comprehensively describing the battery degradation process and accurately predicting the battery health status.

[0070] The present invention proposes a particle filter empirical model for predicting the remaining time of battery fast charging. The model is updated in real time based on the charging conditions (such as current, temperature, SOC, etc.) during the charging process. The model can be adaptively adjusted under different charging conditions to provide accurate fast charging time prediction, thereby optimizing the charging efficiency and user experience of electric vehicles.

[0071] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 Flowchart for predicting remaining battery charging time.

[0073] Figure 2 Schematic diagram of data-driven modeling for battery capacity estimation.

[0074] Figure 3 Schematic diagram of building an empirical model for battery charging time prediction.

[0075] Figure 4 The result diagram of battery capacity estimation.

[0076] Figure 5 This is a graph showing the predicted battery charge capacity.

[0077] Figure 6 Graph showing battery charging time prediction results. DETAILED DESCRIPTION

[0078] like Figure 1 As shown, the present invention mainly includes 4 steps.

[0079] Step 1: Vehicle charging data cleaning and capacity label acquisition. First, use Bluetooth or USB to connect to the vehicle OBD interface to obtain the running and discharging data of the electric vehicle, including time, total voltage, current, SOC (battery state of charge), single cell maximum voltage, single cell minimum voltage, single cell maximum temperature, single cell minimum temperature and other data, record and store the above running data; then, improve the data quality through data cleaning strategies such as threshold screening and nearest neighbor interpolation to ensure the stability and reliability of the data; finally, obtain the battery capacity label based on the fixed SOC interval and variable ampere integral method, which provides key data support for the construction of the model.

[0080] Step 2: Construct graph structure data and extract multi-dimensional spatiotemporal features. First, divide the voltage sequence according to the starting voltage (V_start), ending voltage (V_end) and voltage interval (ΔV), and integrate temperature, capacity and other data to construct multi-dimensional data samples;

[0081]

[0082] Then, the time series of each dimension is used as the node feature vector of the graph structure, and based on the correlation between the nodes, the traditional multi-dimensional time series data is converted into graph structure data. The calculation formula is:

[0083]

[0084] Among them, S pq represents the similarity between the pth and qth nodes; x p and x q are the feature vectors of the pth and qth nodes respectively; j is the total number of nodes;

[0085] The similarities of all nodes are normalized to a uniform scale, and the adjacency matrix A of the graph is expressed as:

[0086]

[0087] Among them, A pq represents the element in the p-th row and q-th column of A.

[0088] Finally, based on the GCN-GRU model framework, the spatiotemporal features of the battery aging process are automatically extracted, and the graph convolution network is used to capture the local features and topological relationships of the battery aging process from the graph structure data; the gated recurrent unit uses the temporal evolution law of the battery aging features extracted and learned from the battery aging process to obtain the key spatiotemporal features of the battery in the entire aging cycle. The calculation process is as follows:

[0089]

[0090] Among them, f a and f c are the aggregation function and update function respectively, N(i) is the node adjacent to i, and σ is the activation function.

[0091]

[0092] Among them, g t and r t are the outputs of the update gate and the reset gate respectively; σ g and σ r They are the activation functions of the update gate and the reset gate respectively; and h represent the candidate hidden state and the final hidden state respectively; W g , W r and W h are the update gate, reset gate and weight matrix of the hidden layer respectively; b g , b r and They are the update gate, reset gate and bias term of the hidden state respectively.

[0093] Step 3: Build a battery capacity prediction model. Based on the acquired label capacity and multi-dimensional graph structure data, a battery capacity prediction model is built using a data-driven approach, such as Figure 2 As shown. First, the model randomly divides the data set into training set and test set by inputting multi-dimensional graph structure samples and capacity label data, and models them in a ratio of 6:4. After initializing the model structure and parameters, the GCN-GRU-based architecture is used to extract the spatiotemporal features of the battery aging process. The feature extraction process is explained in step 2. Then, these spatiotemporal features are mapped to the battery capacity through a fully connected layer. The loss function includes two categories: data loss and physical loss. The model is optimized by minimizing the error between the predicted value and the actual battery capacity. The loss function is shown in the following formula:

[0094]

[0095] Where n is the number of samples, is the capacity value predicted by the model, y is the true capacity, y0 is the initial capacity value, L is the number of cycles, ω1 and ω2 are the weights of data loss and physical loss respectively. The model’s hidden layer parameters and learning rate are continuously updated through the Adam optimizer until the final model parameters are saved after the training cycle ends, and the model’s prediction error and trained parameters are output. Step 4: The charging time prediction process based on the particle filter algorithm is as follows Figure 3 As shown. First, create a particle filter and initialize the state parameters of the capacity degradation empirical formula, then generate initial particles and perform state prediction. The weight of the particle is updated with the new observation value at each moment, and the validity of the weight is ensured by the normalization step. When the particle weight drops to a certain threshold, resampling is performed to supplement the particles with higher weights to ensure the accuracy and robustness of the model. The state prediction and resampling process is executed cyclically until the termination condition is met. Finally, the model parameters are saved and the iteration ends to complete the prediction of the charging time. In actual use, the model is continuously updated based on new real-time data to improve its adaptability and generalization ability to different environments. Whenever new data (such as current, voltage, temperature, SOC, etc.) is input, the particle filter uses the state transition model to make predictions, and uses the observation value to correct the particle weight. By resampling, particles with higher credibility are retained, so that the model can continuously adapt to changes in the battery state, improve the accuracy of the prediction and the generalization ability of multiple scenarios.

[0096] Step 5: The prediction results of the actual vehicle battery capacity based on the GCN-GRU model, the charging amount prediction results based on dynamic particle swarm optimization, and the charging time prediction results are as follows: Figure 4-6 As shown, the battery capacity, charge amount and charging time can be accurately predicted. Figure 4 The comparison between the predicted value and the true value of battery capacity by the GCN-GRU model is shown. The fluctuation trend between the predicted result and the true value is consistent, indicating that the model can effectively capture the changing pattern of battery capacity. Figure 5 The model's ability to predict charging capacity was further verified. The point clouds of predicted values ​​and true values ​​were densely distributed and had a high degree of overlap, indicating that the model had a small error in predicting charging capacity under different charging cycles. Figure 6 The regression relationship between the predicted result of charging time and the true value is shown. The data points are distributed close to the regression line, and most of the predicted values ​​fall within the ±5% error range, indicating that the present invention has high accuracy and reliability in charging time prediction. In general, the method of the present invention has made a breakthrough in solving the problem of fast charging time prediction for electric vehicles by integrating multi-dimensional feature extraction and physical constraints, and also provides important technical support for improving the charging efficiency of new energy vehicles.

Claims

1. A charging time prediction method based on multi-dimensional spatiotemporal characteristics and physical constraints of battery aging, comprising the following steps: S1: Clean vehicle charging data and obtain capacity tags; S11: Obtaining the running and discharging data of the electric vehicle, including: time, total voltage, current, SOC (battery state of charge), single cell maximum voltage, single cell minimum voltage, single cell maximum temperature, single cell minimum temperature, and recording and storing the above running data; S12: When cleaning the running data, first use the threshold range to filter out abnormal data; then use the K-nearest neighbor interpolation method to fill in the missing values ​​to ensure the continuity of the time series data; finally use a low-pass filter to smooth the noise data to ensure the stability and reliability of the data; S13: When obtaining the capacity tag, first set a fixed SOC interval to avoid the influence of extreme sections on the capacity evaluation and ensure that each calculation covers the same battery working state; then use the current and time data after cleaning to calculate the battery capacity based on Formula 1; Among them, C a is the calculated battery capacity, Δt is the sampling interval, I is the charging current, t1 and t2 are the start and end time of discharge respectively, SOC t1 is the discharge start SOC data, SOC t2 The SOC data for discharge termination; S2: Extract multi-dimensional spatiotemporal features; S21: construct multidimensional data samples; The Q(V) curve, Temp(V) curve, Time(V) curve and IC curve are used as inputs of the data-driven model; the number of acquisition points L is determined according to the starting voltage V_start, the ending voltage V_end and the voltage interval ΔV; Wherein, ΔV is set to 2 mV, and the boundary value M of each discharge capacity segment is obtained by setting a fixed length n for each discharge segment; Where c is the step size, and the maximum integer obtained by the function floor(·) does not exceed the input value; In practical applications, it is impossible to obtain the upper limit value of each capacity point; to solve this problem, △Q=[Q1,Q2,…,Q n ]–Q1 is used to replace the capacity sequence, and other temperature and time data are processed in the same way; the final multidimensional data sample is expressed as, Where △Q is the charge change value, V is the voltage value, ΔTemp is the temperature change value, dQ_dV is the IC change value, and n is the length of the collected data; S22: construct a multidimensional data graph structure sample according to the multidimensional data sample; S23: Extracting spatiotemporal features; First, the aggregation function is used in the k convolutional layers of the graph convolutional network to fuse the aggregate information of the neighboring nodes, and then the aggregate information is combined and updated with the current state of the central node. The calculation process is as follows: in, represents the feature vector of node i in the kth layer, represents the feature vector of node j in the k-1th layer, Indicates that j belongs to the node, that is, all nodes, f a and f c are the aggregation function and update function respectively, N(i) is the node adjacent to i, and σ is the activation function; The propagation rule of the kth convolutional layer is: Among them, Z k is the node layer output vector, ω k is the weight matrix; f k (A,X) represents the graph convolution function, X represents the feature matrix, represents the adjacency matrix, X k The feature matrix of the kth layer, ω k represents the weight matrix of the kth layer; The gated recurrent unit is based on Z k As input, the output finally captures the spatiotemporal characteristics of battery aging h t , the calculation formula is: Among them, g t and r t are the outputs of the update gate and the reset gate respectively; σ g and σ r They are the activation functions of the update gate and the reset gate respectively; and h t Represent the candidate hidden state and the final hidden state respectively; W g , W r and W h are the update gate, reset gate and weight matrix of the hidden layer respectively; b g 、b r and They are the bias items of update gate, reset gate and hidden state respectively; S3: Establish a capacity prediction model; S31: Model input data construction and model initialization; When performing capacity prediction based on the GCN-GRU model, GCN stands for graph convolutional network and GRU stands for gated recurrent unit. First, multi-dimensional features are constructed for the input data. The input data is voltage, current, temperature, and time, and graph structure samples are constructed through these features. The constructed graph includes an adjacency matrix and node features. The adjacency matrix represents the correlation between nodes with different features, and the node features represent the specific physical quantity of each node. These graph structure data are passed as input to the graph convolutional network of the model to extract the spatial dependency between features. S32: Spatiotemporal feature extraction of GCN-GRU network; After completing the construction and initialization of the input data, the data is sent to the GCN layer for spatial feature extraction; the GCN layer will process the graph structure data, and extract the correlation between different feature dimensions by aggregating the information of nodes and neighboring nodes; the convolved graph structure features are passed to the GRU network, and the GRU layer is specifically used to extract dynamic features in the time dimension; GRU captures the capacity change trend of the battery at different time points by processing time series data, and extracts the time series features of the battery aging process; finally, through the fully connected layer, these spatiotemporal features are mapped to the output value of the battery capacity estimation; S33: Determine physical constraints on aging rate and model optimization; During the model training process, the loss function includes two types: data loss and physical loss. The model is optimized by minimizing the error between the predicted value and the actual battery capacity. The loss function is shown as follows: Among them, Loss data Represents the data loss function, Loss physics represents the physical feature loss function, Loss represents the total loss, ω1 and ω2 represent the proportional coefficients, and L i represents the number of historical cycles of the i-th data, n is the number of samples, is the capacity value predicted by the model, y i is the true capacity, y0 is the initial capacity value, L i is the number of cycles; The Adam optimizer is used in training to accelerate the convergence of the model with an adaptive learning rate, so that it can better fit the changing trend of battery capacity. In order to prevent overfitting, the data set is divided into a training set and a validation set. The model performance is monitored and hyperparameters are adjusted through the validation set to ensure the generalization ability of the model under different charging and discharging conditions, and finally obtain stable and accurate capacity prediction results. S4: Predict charging time based on particle filter algorithm; S41: Empirical model construction and parameter initialization; In the empirical model of charging capacity prediction, the initial model is first constructed based on the particle filter algorithm, and the parameters are initialized; the model input includes battery capacity, charging SOC range, charging temperature range and charging rate range; the state evolution of these variables is simulated by the particle filter, and the charging process model in multi-dimensional space is constructed; The goal of the model is to dynamically update and predict the hidden variables of the charging process through the particle filtering process; the initialization step includes selecting a set of particles to represent the state distribution and estimating the charging time based on the state transition model; the initialization formula of the particle filtering algorithm is: p(x0)~N(μ0,Σ0) Among them, x0 is the initial state variable; μ0 and Σ0 are the mean and variance of the initial state; p(x0) represents the probability distribution of the initial state x0, N(μ0,∑0) represents the normal Gaussian distribution with mean μ0 and covariance matrix ∑0, and ~ represents "obey" or "comply". S42: Continuously update the model based on new input data; At each time step, the particle filter corrects the weight and state of the particle through the state transfer equation and the observation equation, and finally estimates the actual charging time more accurately. The observation data is continuously input into the model, and the particle filter resamples the particle weight and estimates the actual charging amount in combination with the new observation value. The particle filter formula for state update and observation is: p(x t |y 1:t )∝p(y t |x t )∫p(x t |x t-1 )p(x t-1 |y 1:t-1 )dx t-1 Among them, p(x t y 1:t ) means that all observations y from time 1 to t are obtained 1：t After that, for the state x at time t t The posterior probability distribution of y 1:t represents the set of all observations from time 1 to time t, that is, {y1, y2, ..., y t }, ∝ means proportional to, x t is the state variable at the tth moment, y t is the observed value at the current moment, p(y t |x t ) is the observation model, which represents the state x t Calculate the observation y t The probability of p(x t |x t-1 ) is a state transition model, which indicates the transition probability of the current state based on the previous state.

2. A charging time prediction method based on multi-dimensional spatiotemporal characteristics and physical constraints of battery aging as claimed in claim 1, characterized in that: The specific method of step S22 is: The multi-dimensional time series data is converted into graph structure data. In this process, the time series of each dimension is used as the node feature vector of the graph structure. The edge construction method is based on the correlation between the nodes, and the calculation formula is: Among them, S pq represents the similarity between the pth and qth nodes; x p and x q are the eigenvectors of the pth and qth nodes respectively; j is the total number of nodes, ||·||2 is the matrix bi-norm; The similarities of all nodes are normalized to a uniform scale; the adjacency matrix A of the graph is expressed as: Among them, A pq represents the element in the p-th row and q-th column of A.

3. The charging time prediction method based on multi-dimensional spatiotemporal characteristics and physical constraints of battery aging according to claim 1, characterized in that: In step S23, the features are further screened. The screening method is: when the node feature dimension is greater than the set threshold, a portion of elements on the feature vector are randomly shielded to 0, and the discarded ratio is determined by the hyperparameter p; the features after discarding are calculated as follows: Among them, Z k ' is the feature matrix after the Dropout operation; Dropout(·, p) is the Dropout operation with a discard ratio of p.

4. The charging time prediction method based on multi-dimensional spatiotemporal characteristics and physical constraints of battery aging according to claim 1, characterized in that: The calculation formula of the optimizer in step S33 is as follows: Where η is the learning rate, θ t is the current model parameter, m t is the momentum of the gradient, v t is the weighted average of the squared gradients, and ∈ is a small constant to prevent division by zero.

Images