New energy automobile battery recycling management method

Abstract

The invention discloses a new energy vehicle battery recycling management method, and relates to the field of new energy vehicles, and the method comprises the steps: collecting multi-source heterogeneous data of a vehicle; the preprocessed historical vehicle track and the preprocessed battery parameters are regarded as different views respectively, and corresponding features are extracted: the features of the historical vehicle track view and the features of the battery parameter view are fused by adopting a multi-view subspace clustering algorithm to obtain vehicle-battery comprehensive features; calculating the similarity between the comprehensive characteristics of different vehicles-batteries, and constructing a hierarchical clustering tree of the vehicles-batteries to obtain a clustering result of the vehicles-batteries; in each cluster, a decision tree algorithm is adopted to train a battery residual capacity prediction model; and generating a battery recovery strategy of each vehicle-battery cluster according to a vehicle-battery clustering result and a prediction result of the residual capacity of the batteries in the corresponding cluster. Aiming at the lack of pertinence of new energy automobile battery recovery in the prior art, the battery recovery accuracy is improved.

Description

Technical Field

[0001] The present application relates to the field of new energy vehicles, and in particular to a new energy vehicle battery recycling and utilization management method. Background Art

[0002] As global energy shortages and environmental pollution become increasingly severe, the development of new energy vehicles has become an important direction for the transformation and upgrading of the automotive industry. As a core component of new energy vehicles, the service life and recycling efficiency of power batteries are crucial to the sustainable development of the new energy vehicle industry. However, the current recycling of new energy vehicle batteries faces many challenges, and it is urgent to develop accurate and efficient battery recycling management technologies to extend battery life, improve battery recycling rates, and promote the healthy development of the new energy vehicle industry.

[0003] However, most existing battery recycling strategies adopt a "one-size-fits-all" approach, that is, determining a unified recycling time node based on simple indicators such as battery usage time or mileage, ignoring the differences between different vehicle-battery individuals in terms of usage conditions, driving modes and health status, resulting in a lack of targeted recycling strategies for new energy vehicle batteries. Summary of the invention

[0004] In response to the problem of lack of targeted recycling strategies for new energy vehicle battery recycling in the prior art, the present application provides a new energy vehicle battery recycling and utilization management method. Through multi-view feature extraction and fusion, on the basis of hierarchical clustering analysis, a decision tree algorithm is used to accurately predict the remaining battery capacity in each cluster, and differentiated battery recycling strategies are generated according to the clustering results and the battery remaining capacity prediction results, thereby improving the accuracy of new energy vehicle battery recycling.

[0005] The purpose of this application is achieved through the following technical solutions.

[0006] The present application provides a new energy vehicle battery recycling management method, including: collecting multi-source heterogeneous data of a vehicle, the multi-source heterogeneous data including vehicle history trajectory data and battery parameters; wherein the battery parameters include SOC data; the vehicle history trajectory data includes GPS positioning data; preprocessing the collected multi-source heterogeneous data; treating the preprocessed vehicle history trajectory and battery parameters as different views, and extracting corresponding features: extracting the vehicle's driving trajectory features as vehicle history trajectory view features; extracting the battery's charging and discharging features as battery parameter view features; using a multi-view subspace clustering algorithm to fuse the vehicle history trajectory view features and the battery parameter view features to obtain to the comprehensive characteristics of the vehicle-battery; based on the comprehensive characteristics of the vehicle-battery, the similarities between the comprehensive characteristics of different vehicle-batteries are calculated, and a hierarchical clustering tree of the vehicle-battery is constructed; according to the hierarchical clustering tree, the clustering results of the vehicle-battery are obtained; according to the clustering results of the vehicle-battery, in each cluster, the vehicle historical trajectory data and battery parameter data are used as training samples, and a decision tree algorithm is used to train a battery remaining capacity prediction model; the trained battery remaining capacity prediction model is used to predict the battery remaining capacity of the vehicles in the cluster; according to the clustering results of the vehicle-battery and the prediction results of the battery remaining capacity in the corresponding cluster, a battery recycling strategy for each vehicle-battery cluster is generated.

[0007] Furthermore, the collected multi-source heterogeneous data is preprocessed, including: using the Kalman filter algorithm to denoise the GPS positioning data in the vehicle historical trajectory data to eliminate GPS signal errors and drifts. The GPS latitude and longitude, speed and other information in the vehicle historical trajectory data are used as the observation quantity of the Kalman filter. The state transfer equation and observation equation of the Kalman filter are established based on the vehicle kinematic model. Through the prediction and update steps of the Kalman filter, the real position and speed state of the vehicle are recursively estimated to eliminate the Gaussian white noise and deviation errors in the GPS positioning data. The B-spline curve fitting algorithm is used to smooth the vehicle historical trajectory data after Kalman filtering to remove mutations in the trajectory data. For each vehicle's historical trajectory sequence, a set of time points are selected as the control points of the B-spline curve. Based on the control points, the cubic B-spline curve is fitted by the least squares method to generate a smooth trajectory curve. The B-spline curve has piecewise continuous smoothness and can effectively remove abnormal fluctuations in trajectory data caused by GPS signal mutations, road turns, etc.

[0008] According to the battery model and capacity parameters, the SOC (state of charge) data of different batteries are normalized. For batteries of different models and capacity specifications, their SOC ranges may be different and need to be uniformly mapped. Suppose the rated capacity of the i-th battery is Q i , the lower limit of SOC is soc mini , the upper limit of SOC is socmaxi The original SOC value soc collected at any time for the i-th battery i , is mapped to the interval [0, 1] by the following formula: After normalization, the SOC data of different batteries are unified to the same scale to facilitate subsequent cluster analysis. The local anomaly factor LOF algorithm is used to detect anomalies in the normalized battery parameters. For each battery, its SOC, voltage, current, temperature and other monitoring parameters form a feature vector at each sampling time. The LOF algorithm is used to measure the degree of local density deviation of each feature vector relative to its surrounding neighborhood. The larger the LOF value, the lower the density of the vector point relative to the neighboring points, and the greater the possibility of anomaly. Set a LOF threshold and mark the sampling time where the feature vector whose LOF value exceeds the threshold is located as an anomaly point. The number and proportion of abnormal points in each battery monitoring parameter sequence are counted as the basis for whether the battery is abnormal.

[0009] Furthermore, the vehicle's driving trajectory features are extracted as the vehicle's historical trajectory view features, including: calculating the distance between adjacent trajectory points based on the GPS positioning data in the vehicle's historical trajectory data, obtaining the mileage of each trajectory segment, and generating a mileage sequence. For each vehicle's GPS trajectory point sequence, two adjacent trajectory points form a trajectory segment. Using the conversion formula from longitude and latitude coordinates to spherical distance, the spherical distance of each trajectory segment is calculated as the mileage of the trajectory segment. The mileage of each trajectory segment is organized into a mileage sequence {d1, d2, ..., d n Set sliding windows of different sizes, perform sliding aggregation on the mileage sequence, and sum the mileage in each sliding window to obtain a multi-scale mileage feature sequence. Set the size of the sliding window to {w1,w2,.....,w m}, where w1<w2<.....<w m For scale w i , sliding size w on the mileage sequence with a step size of 1 i The window is used to calculate the mileage accumulation in each window to form a scale w i The mileage feature sequence {s i (1),s i (2),.....,s i (nw i +1)}. By setting sliding windows of different scales, the mileage variation patterns of vehicles in different time spans can be captured.

[0010] Using the statistical feature extraction algorithm, the mileage feature sequence at each scale is extracted to obtain the mileage statistical features at the corresponding scale.i The mileage characteristic sequence under

[0011] {s i (1),s i (2),.....,s i (nw i +1)}, calculate its mean u i , variance v i and kurtosis k i :

[0012]

[0013] The scale w is composed of mean, variance and kurtosis i The mileage statistical feature vector {u i ,v i ,k i}. Traverse sliding windows of different scales to obtain a set of multi-scale mileage statistical features. For m different sliding window scales {w1,w2,.....,w m}, respectively perform the feature extraction steps of S32 and S33. Finally, a set of multi-scale mileage statistical feature vectors is obtained

[0014] {{u1,v1,k1},{u2,v2,k2},......,{u m ,v m ,k m The feature cascade algorithm is used to cascade the multi-scale mileage statistical features in scale order to generate the vehicle history trajectory view features. The mileage statistical feature vectors extracted at different scales are spliced ​​in order from small to large scale to form a 3m-dimensional feature vector: {u1,v1,k1,u2,v2,k2,.....,u m ,v m ,k m}, the above feature vector is the extracted vehicle history trajectory view feature, denoted by X traj .

[0015] Furthermore, the battery charge and discharge characteristics are extracted as the battery parameter view features, including: according to the preprocessed battery parameter data, the battery SOC data is divided into charge and discharge cycles using a cycle detection algorithm to obtain each complete charge and discharge cycle data. The SOC time series data is analyzed to detect the maximum and minimum points of the SOC, and the charge and discharge state is determined according to the threshold conditions. A complete charge and discharge cycle from a maximum point to the next maximum point is taken as a complete charge and discharge cycle, and the SOC subsequence data within each cycle is extracted. The battery SOC charge and discharge cycle set {C1, C2, ....., C k}, where C iIndicates the SOC data of the i-th charge and discharge cycle.

[0016] For each charge and discharge cycle data, the SOC data is resampled using an interpolation algorithm to obtain an SOC curve with the same length. Since the time span of different charge and discharge cycles may be different, the SOC data needs to be normalized for comparison. A fixed resampling length L is set, and the SOC sequence C of each cycle is i Perform linear interpolation or spline interpolation to obtain a resampled SOC sequence C of length L i '. The resampled SOC sequence C i ' can be regarded as the SOC curve on the normalized time scale, and the length of the SOC curve in different cycles is consistent. The SOC curve is fitted with a polynomial function, and the SOC range, SOC change rate and cumulative discharge amount are extracted as a set of characteristic vectors of the corresponding cycle charge and discharge characteristics. For the resampled SOC curve C of the i-th charge and discharge cycle i ', use n-order polynomial function for fitting: soc = a0 + a1t + a2t 2 +.....+a n t n , where t is the normalized time variable, t∈[0,1]. The polynomial coefficients a0 to a n The changing trend of the SOC curve is depicted.

[0017] Based on the fitting curve, the following features are extracted: SOC range: soc range_i =max(C i ')-min(C i '); Average SOC change rate: Cumulative discharge capacity:

[0018] By {soc range_i ,soc rate_i ,soc discharge_i} constitutes the charge and discharge characteristic vector F of the i-th cycle i . Extract the feature vectors of all charge and discharge cycles and arrange them in order to obtain the time series of battery charge and discharge features {F1, F2, ..., F k The sliding window method is used to calculate the mean, variance and skewness of each feature at different time scales, and multi-scale battery charge and discharge statistical features are generated as battery parameter view features. Set sliding windows of different sizes {w1,w2,......,w m}, sliding on the charge and discharge feature time series. For feature j (j = 1, 2, 3) and window size w i , calculate the corresponding mean u ij , variance vij and skewness s ij : Finally, we get the battery parameter view feature X bat , which contains the mean, variance, and skewness of each charge and discharge feature at all sliding window scales.

[0020] Furthermore, a multi-view subspace clustering algorithm is used to fuse the vehicle history trajectory view features and the battery parameter view features to obtain the comprehensive features of the vehicle-battery, including: using the principal component analysis algorithm PCA to reduce the dimension of the original features of the vehicle history trajectory view to obtain the compressed vehicle history trajectory view features V1. For each vehicle, its historical trajectory GPS sequence, speed sequence, etc. constitute the original feature matrix X1 of the trajectory view. After X1 is centralized, the principal component vector is obtained by eigenvalue decomposition, and the first k1 principal component vectors are taken to form the PCA projection matrix P1. X1 is mapped to the k1-dimensional subspace defined by P1 to obtain the compressed trajectory view features V1.

[0021] The independent component analysis algorithm ICA is used to perform dimensionality reduction mapping on the original features of the battery parameter view to obtain the compressed battery parameter view features V2. For each vehicle, its battery capacity, voltage and other monitoring parameters constitute the battery view original feature matrix X2. After centering and whitening X2, the independent component vector is estimated by the FastICA algorithm, and the first k2 independent component vectors are taken to form the ICA projection matrix P2. X2 is mapped to the k2-dimensional subspace defined by P2 to obtain the compressed battery view features V2. The cosine similarity between the compressed vehicle history trajectory view features V1 and the battery parameter view features V2 is calculated to obtain the inter-view similarity matrix S, which is used to reflect the correlation between different view features. For V1 and V2 of the i-th vehicle, their cosine similarity S is calculated. i =cos(V 1i ,V 2i ). The similarity matrix S is constructed by the cosine similarity of all vehicles, where S ij represents the similarity of the compressed features of the i-th vehicle in two views.

[0022] The following multi-view reconstruction error objective function is constructed, using the compressed vehicle history trajectory view feature V1 to reconstruct X1, and using the compressed battery parameter view feature V2 to reconstruct X2: Among them, a and b are the reconstruction error weights of the two views. Based on the multi-view reconstruction error objective function, the similarity matrix S between views is introduced, and the regularization term is constructed to generate the following overall optimization objective function for multi-view subspace representation learning:

[0023] Where: V is the n×(k1+k2)-dimensional joint compressed feature matrix formed by concatenating V1 and V2 by row, and D is the diagonal matrix of S, that is, D ii =∑j×S ij , tr(*) represents the trace of the matrix, and c is the weight coefficient of the regularization term. The regularization term is used to minimize the graph Laplacian operator of the compressed features, so that the samples with high similarity between views should have closer learned compressed features.

[0024] The overall optimization objective function of S35 is solved by an alternating optimization strategy, that is, V and P are fixed alternately, and the other parameter is optimized respectively until the objective function converges. When P is fixed, the subproblem of optimizing V is a quadratic function, and a closed-form solution can be obtained by differentiation. When V is fixed, the subproblem of optimizing P becomes a classic ridge regression problem, which can also be solved explicitly. The optimization of the above two subproblems is iteratively performed on each vehicle until the overall objective function converges, and V1 and V2 of the current vehicle are obtained. Finally, the optimized compressed features V1 and V2 of all vehicles are obtained, and they are spliced ​​by row to form a comprehensive vehicle-battery feature representation.

[0025] Further, according to the hierarchical clustering tree, the clustering results of the vehicle-battery are obtained, including: according to the vehicle-battery comprehensive features obtained by clustering, the similarity between different vehicle-battery comprehensive features is calculated using the Pearson correlation coefficient to obtain a vehicle-battery similarity matrix; generating a hierarchical clustering tree, including: initializing each vehicle-battery comprehensive feature vector in the vehicle-battery similarity matrix as an independent cluster to obtain an initial cluster set. If there are n vehicle-battery pairs, the vehicle-battery similarity matrix is ​​an n×n symmetric matrix D, where D ij Represents the similarity between the i-th vehicle-battery pair and the j-th vehicle-battery pair. Initialize the cluster set C0 = {{1}, {2}, ..., {n}}, where {i} represents the cluster containing only the i-th vehicle-battery pair.

[0026] The single connection minimum distance algorithm is used to calculate the distance between each independent cluster in the initial cluster set and other clusters in the vehicle-battery similarity matrix, and the two clusters with the smallest distance value are merged into a new cluster. For any two clusters A and B in the cluster set C_t, their single connection distance is defined as: d(A,B)=min{i∈A,j∈B}D ij , in C t Find the two clusters A and B with the smallest single connection distance and merge them into a new cluster M = A*∪B*. Update the cluster set C t+1 =C t{A*,B*}∪{M}. Based on the merged new cluster, update the vehicle-battery comprehensive feature similarity matrix, and use the single connection minimum distance algorithm to recalculate the distance between the new cluster and other clusters. For the newly merged cluster M, calculate its distance from C t+1 The single connection distance of other clusters in the cluster is used to update the similarity matrix D:

[0027] Represents C t+1 The elements belonging to M are excluded, S424, the step of merging clusters is repeated until only one cluster remains in the current cluster set, and a hierarchical clustering tree is generated. S422 and S423 are repeated until only one element remains in the cluster set, namely C T ={1,2,......,n}. The whole cluster merging process generates a series of nested cluster sets {C1,C2,.....,C T}, reflecting the hierarchical structure of clustering.

[0028] All clusters generated in the cluster merging process are used as nodes of the hierarchical clustering tree, among which the initial independent clusters are used as leaf nodes, the clusters generated by the intermediate merging are used as internal nodes, and the result of the final merging into one cluster is used as the root node to obtain a hierarchical clustering tree. Each independent cluster {i} in the initial cluster C0 is used as a leaf node of the hierarchical clustering tree. For each new cluster M generated by cluster merging, it is added to the hierarchical clustering tree as an internal node, and the two merged clusters A and B are used as child nodes of M. The final cluster containing all vehicle-battery pairs is used as the root node of the hierarchical clustering tree. The tree structure composed of leaf nodes, internal nodes and root nodes is the hierarchical clustering tree, denoted as T. Set the global fitness function, and use the genetic algorithm to search for the global optimal partition of the hierarchical clustering tree to obtain the clustering results of the comprehensive characteristics of the vehicle-battery;

[0029] Furthermore, a genetic algorithm is used to search for the global optimal partition of the hierarchical clustering tree, including: setting the global fitness function as the ratio of the sum of the similarity values ​​of the vehicle-battery comprehensive characteristics within a cluster to the sum of the similarity values ​​of the vehicle-battery comprehensive characteristics between clusters. For a partitioning scheme P of the hierarchical clustering tree T, k clusters {C1, C2, ....., C k}. Define the intra-cluster similarity S in and the similarity between clusters S out : The global fitness function is defined as: That is, the ratio of the similarity within a cluster to the similarity between clusters.

[0030] Randomly select the partition point of the vehicle-battery comprehensive feature hierarchical clustering tree to generate the initial partition population, and each partition individual corresponds to a clustering result. Randomly select m internal nodes in the hierarchical clustering tree T as partition points, and regard these m nodes and their subtrees as independent clusters. The rest of T is regarded as another cluster to form a partition individual. Repeat the above process n times to generate an initial population P0 containing n partition individuals. For each partition individual in the initial partition population, calculate the corresponding global fitness function value as the individual fitness. For the population P t Each partition individual P in i , calculate its fitness value f(P i ).

[0031] The genetic algorithm is used to iterate the current population until the maximum number of iterations is reached or the average fitness of the population reaches the preset threshold. t , perform the following genetic algorithm operations: Selection: According to the individual fitness value, use methods such as roulette or tournament selection to select from P t Select a certain number of excellent individuals from P. Crossover: Perform a crossover operation on the selected excellent individuals, randomly exchange some of their partitioning points, and generate new partitioning individuals. Mutation: Mutate the newly generated partitioning individuals with a certain probability, and randomly change some of the partitioning points. Update: Replace the newly generated partitioning individuals with P. t The individuals with lower fitness in the population are obtained as the new population P t+1 The above process is repeated until the maximum number of iterations is reached or the average fitness of the population reaches the preset threshold. From the final population, select the individual with the highest global fitness, and use the corresponding clustering result as the global optimal partitioning result of the hierarchical clustering tree to obtain the clustering result of the vehicle-battery comprehensive characteristics. After the genetic algorithm iteration is completed, t Select the partition individual P* with the largest global fitness function value. The clustering results corresponding to P {C1, C2*, ....., C k *} is the global optimal partition of the hierarchical clustering tree T, which is also the final clustering result of the vehicle-battery comprehensive characteristics.

[0032] Furthermore, a decision tree algorithm is used to train a battery remaining capacity prediction model; including: extracting the historical trajectory data of the vehicle and the battery parameter data at the corresponding time in each cluster to construct a training sample set; for each cluster C in the clustering result i , extract the historical trajectory data and battery parameter data of all vehicles in the cluster. For each vehicle, its historical trajectory data is matched with the battery parameter data collected at the same time to form a set of feature-label samples (x ij ,y ij). Feature x ij Including vehicle historical trajectory characteristics (such as average speed, acceleration, fuel consumption, etc.) and battery parameter characteristics (such as voltage, current, temperature, etc.). ij is the remaining capacity of the battery at that moment. i The feature-label samples of all vehicles in the cluster are summarized to form the training sample set D of the cluster. i ={(x ij ,y ij ),j=1,2,.....n i The CART decision tree algorithm is used to train the battery remaining capacity prediction model. The training sample set is used as the model input, and the battery remaining capacity is used as the target variable. The decision tree is recursively generated through feature selection and sample partitioning. The maximum depth and the minimum number of node samples of the CART decision tree are optimized using the cross-validation method to obtain the optimized battery remaining capacity prediction model. The optimized battery remaining capacity prediction model is used to predict the battery remaining capacity of the vehicles in the cluster.

[0033] Furthermore, the CART decision tree algorithm is used to train the battery remaining capacity prediction model, including: constructing a CART decision tree model, taking the training sample set as the model input, taking the battery remaining capacity as the target variable, and recursively generating a decision tree through feature selection and sample partitioning. Initialize the root node of the decision tree, and take the training sample set D i Input to the root node. Recursively perform feature selection and sample partitioning on each node to generate a decision tree. Use the Gini index as the evaluation criterion to calculate the impurity of each feature for sample partitioning, select the feature with the smallest Gini index as the optimal split feature, and perform binary tree splitting on the current node sample set according to the value of the optimal split feature to form left and right child nodes. For node t, its sample set is D t , calculate each feature A k The Gini index is: in, For feature A k The sample subset with value v, is the Gini impurity of the subset. Select feature A with the smallest Gini index k * as the optimal splitting feature. According to A k *The value of D t Divided into sample sets D1 and D2 of left and right child nodes r . Recursively split the left and right child nodes until one of the decision tree growth stopping conditions is met. The decision tree growth stopping conditions include: the number of samples at the current node | D t| is less than the preset sample threshold. The samples of the current node have the same value on the target variable y. The feature values ​​of the current nodes are the same. The depth of the decision tree reaches the preset maximum depth. The minimum Gini index of the current node is greater than or equal to the Gini index of the parent node.

[0034] The nodes that meet the decision tree growth stop condition are marked as leaf nodes, and the average remaining battery capacity of the corresponding samples is used as the output value of the corresponding leaf node. For leaf node t, its output value is: Prune the leaf nodes, calculate the loss function before and after pruning for each leaf node from bottom to top, and when the reduction value of the loss function after pruning is less than the threshold, prune the corresponding leaf node and the corresponding parent node and merge them into a new leaf node until the reduction value of the loss function after pruning of all leaf nodes is greater than the threshold, and obtain the initial CART decision tree model. Use the cross-validation method and use a separate validation set to evaluate the pruning effect of each leaf node. Define the change in the loss function before and after pruning: Among them, R(t) is the error sum of leaf node t, R(T t ) is the subtree T with t as the root t The sum of the squared errors of the leaf nodes, |T t | is the number of leaf nodes in the subtree. For leaf nodes whose ΔL(t) is less than the preset threshold, merge them with the corresponding parent node into a new leaf node and update the decision tree structure. Repeat the pruning process until all ΔL(t) are greater than the threshold, and obtain the final CART decision tree model.

[0035] Furthermore, the maximum depth of the decision tree and the minimum number of samples at the node are optimized by the cross-validation method to obtain the optimized battery remaining capacity prediction model, including: using grid search to determine the candidate value range of the hyperparameter, the hyperparameter includes the maximum depth of the CART decision tree and the minimum number of samples at the leaf node. The candidate value range of the maximum depth of the decision tree is set to [d min ,d max ], with an interval of Δd. Set the candidate value range of the minimum number of samples of leaf nodes to [n min ,n max ], with an interval of Δn. Generate a hyperparameter grid, where each grid point corresponds to a set of hyperparameter candidate value combinations (d,n). Divide the training sample set into K mutually exclusive subsets, select K-1 subsets as training sets each time, and the remaining 1 subset as validation set. Use the K-fold cross-validation method to divide the training sample set D i Randomly divide into K mutually exclusive subsets In each cross-validation, K-1 subsets are selected to form the training set D train , the remaining subset is used as the validation set D valFor each set of candidate hyperparameter values, train and evaluate the performance of the CART decision tree model under the current parameter combination on the validation set. train is the training set, and the CART decision tree model M(d,n) is trained according to the steps in S52. val The prediction performance of the model M(d,n) is evaluated above, and the root mean square error (RMSE) is used as the evaluation indicator: in, For the model to sample x j The prediction output of . Record the cross-validation performance RMSE under the current hyperparameter combination (d, n) k (d,n).

[0036] Traverse all hyperparameter candidate value combinations, select the hyperparameter combination with the best model performance, and obtain the optimized CART decision tree model as the battery remaining capacity prediction model. For each set of hyperparameter candidate value combinations (d, n), calculate the average performance of its K-fold cross validation: Select the hyperparameter combination (d*,n*) with the best average performance: Take (d*,n*) as the hyperparameter and use the entire training sample set D i The optimized CART decision tree model M* is trained and used as the battery remaining capacity prediction model within the cluster.

[0037] Furthermore, based on the clustering results of vehicles and batteries and the prediction results of the remaining battery capacity in the corresponding cluster, a battery recycling strategy for each vehicle-battery cluster is generated, including: based on the predicted distribution of the remaining battery capacity of the vehicles in each cluster and combined with the preset remaining battery capacity threshold, the battery recycling time window corresponding to each cluster is determined. i , obtain the predicted remaining capacity of all vehicle batteries in the cluster in the future time period {Q ij (t),j=1,2,.....,n i}. Calculate cluster C based on the predicted remaining capacity i Distribution of battery remaining capacity P in each time period i (Q,t). Set one or more battery remaining capacity thresholds {Q1,Q2,......,Q m}, represents the remaining capacity state of the battery when it needs to be recycled. For each threshold Q k , according to cluster C i The remaining battery capacity distribution P i (Q, t), calculate the remaining battery capacity in the cluster to reduce to Q k The time required t ik According to {t ik,k=1,2,......,m} determine the cluster C i Battery recycling time window Indicates that the vehicle batteries in this cluster need to be recycled within this time frame.

[0038] Based on the historical driving trajectory distribution of each vehicle and the spatial distribution results of the global vehicle-battery clusters, differentiated battery recycling outlets are planned and deployed for different vehicle-battery clusters. j Historical driving GPS trajectory data {(x jt ,y jt ),t=1,2,......,T j}, and obtain the historical driving trajectory distribution of each vehicle. For each vehicle-battery cluster C i , calculate the spatial aggregation area A of the vehicle trajectory distribution within the cluster i , indicating the main area of ​​vehicle activity in this cluster. Comprehensively considering the spatial aggregation area of ​​each cluster {A i ,i=1,2,.....,k} and the spatial distribution of vehicle trajectories in the entire region, plan the location and number of battery recycling outlets. In areas where vehicle trajectories are dense and multiple clustering spatial aggregation areas overlap, deploy key recycling outlets. In areas where vehicle trajectories are relatively scattered and different clustering spatial aggregation areas are independent, deploy scattered recycling outlets. For each cluster C i , according to its battery recycling time window Spatial Aggregation Area A i , vehicle trajectory distribution and other factors to develop differentiated battery recycling plans. Determine the recycling time of the vehicle batteries in the cluster. Select a suitable time point within the range. Select the location where the cluster vehicle trajectory distribution has a high match with the recycling network as the recycling location to reduce the vehicle battery transportation cost. Combined with the battery health status, predicted remaining capacity and other characteristics of the cluster, formulate targeted battery recycling and cascade utilization strategies.

[0039] Compared with the prior art, the advantages of this application are:

[0040] By extracting the mileage statistical characteristics and battery charging and discharging characteristics of the vehicle's historical trajectory and using a multi-view subspace clustering algorithm for feature fusion, the comprehensive characteristics of the vehicle-battery can be characterized from the two perspectives of driving mode and battery usage mode, fully reflecting the intrinsic connection between vehicle usage conditions and battery health status.

[0041] A hierarchical clustering tree is constructed based on the comprehensive characteristics of vehicles and batteries, and a genetic algorithm is introduced to search for the global optimal partitioning. This allows for fine-grained clustering of vehicles and batteries based on comprehensive consideration of similarities within clusters and differences between clusters. This ensures that vehicles and batteries within the same cluster have high consistency in driving modes and battery usage patterns, thus laying the foundation for developing targeted battery recycling strategies.

[0042] In each vehicle-battery cluster, the vehicle historical trajectory and battery parameter data are used as training samples, and the CART decision tree algorithm is adopted to train the battery remaining capacity prediction model. The model hyperparameters are optimized through cross-validation. This can fully explore the personalized characteristics of the vehicle-battery in the cluster, improve the accuracy of the battery remaining capacity prediction, and provide a quantitative basis for formulating the optimal battery recycling time window.

[0043] According to the vehicle-battery clustering results and the battery remaining capacity prediction results, the battery health status distribution and degradation trend of each cluster are analyzed to determine the differentiated battery recycling time window to maximize the battery life and improve battery utilization. By preprocessing the vehicle historical trajectory data and battery parameter data, including Kalman filter denoising, B-spline curve smoothing, SOC data normalization and LOF anomaly detection, the noise interference and abnormal deviation in the data can be effectively eliminated to improve the data quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is an exemplary flow chart of a new energy vehicle battery recycling and utilization management method shown in this application;

[0045] Figure 2 is an exemplary flow chart for preprocessing multi-source heterogeneous data shown in the present application;

[0046] Figure 3 is an exemplary flow chart of generating a vehicle history track view feature as shown in the present application;

[0047] Figure 4 is an exemplary flow chart of generating a battery parameter view feature as shown in the present application;

[0048] Figure 5 is an exemplary flow chart for obtaining vehicle-battery comprehensive characteristics shown in the present application;

[0049] Figure 6 is an exemplary flow chart of obtaining clustering results of vehicle-battery comprehensive features as shown in the present application;

[0050] Figure 7 It is an exemplary flowchart of constructing a battery remaining capacity prediction model shown in this application. DETAILED DESCRIPTION

[0051] The method and system provided in the embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0052] Figure 1 It is an exemplary flow chart of a new energy vehicle battery recycling management method shown in the present application, comprising: collecting multi-source heterogeneous data of the vehicle, the multi-source heterogeneous data including vehicle historical trajectory data and battery parameters; wherein the battery parameters include SOC data; the vehicle historical trajectory data includes GPS positioning data; preprocessing the collected multi-source heterogeneous data; treating the preprocessed vehicle historical trajectory and battery parameters as different views, and extracting corresponding features: extracting the vehicle's driving trajectory features as vehicle historical trajectory view features; extracting the battery's charging and discharging features as battery parameter view features; using a multi-view subspace clustering algorithm to cluster the vehicle historical trajectory view features and the battery parameter view features Fusion, to obtain the comprehensive characteristics of vehicle-battery; based on the comprehensive characteristics of vehicle-battery, calculate the similarity between the comprehensive characteristics of different vehicle-batteries, and construct a hierarchical clustering tree of vehicle-battery; according to the hierarchical clustering tree, obtain the clustering results of vehicle-battery; according to the clustering results of vehicle-battery, in each cluster, use the vehicle historical trajectory data and battery parameter data as training samples, and use the decision tree algorithm to train the battery remaining capacity prediction model; use the trained battery remaining capacity prediction model to predict the battery remaining capacity of vehicles in the cluster; according to the clustering results of vehicle-battery and the prediction results of the battery remaining capacity in the corresponding cluster, generate a battery recycling strategy for each vehicle-battery cluster.

[0053] Collect multi-source heterogeneous data of the vehicle, including vehicle historical trajectory data and battery parameters; battery parameters include SOC data; vehicle historical trajectory data includes GPS positioning data; specifically, data acquisition equipment: vehicle-mounted GPS positioning device, data content: including latitude and longitude, altitude, speed, direction and other information; data acquisition frequency: once every 30 seconds; data transmission method: through the vehicle-mounted 4G network module, the collected GPS data is uploaded to the cloud server for storage in real time. Battery parameter data collection, data acquisition equipment: vehicle-mounted battery management system (BMS); data content: including battery SOC, voltage, current, temperature and other parameters; data acquisition frequency: once every minute; data transmission method: through the vehicle-mounted CAN bus, the data collected by BMS is transmitted to the vehicle-mounted central control system, and then uploaded to the cloud through the 4G network.

[0054] Figure 2This is an exemplary flow chart of preprocessing multi-source heterogeneous data shown in the present application, preprocessing the collected multi-source heterogeneous data, including: using the Kalman filter algorithm to denoise the GPS positioning data in the vehicle historical trajectory data to eliminate GPS signal errors and drifts. Specifically, the state variable vector is defined as: X k =[x k ,y k ,v x,k ,v y,k ] T , where x k and k They represent the longitude and latitude coordinates of the vehicle at time k, v x,k and v y,k Represent the velocity components of the vehicle in the x and y directions respectively. The observation variable vector is defined as in and They represent the longitude and latitude coordinates measured by the GPS device at time k. Establish the state transition equation: X k =FX k-1 +w k , where F is the state transfer matrix, w k is process noise, with mean zero and covariance matrix Q k The multivariate Gaussian distribution, w k ~N(0,Q k ). Establish the observation equation: Z k =HX k +v k , where H is the observation matrix, v k is the observation noise, with a mean of zero and a covariance matrix of R k Multivariate Gaussian distribution of . Set the initial state estimate And its covariance matrix P0. Recursive filtering, prediction step: According to the state estimate and state transfer equation at the previous moment, predict the state prior estimate at the current moment: Update step: Combine the current observations and calculate the Kalman gain to update the state posterior estimate and its covariance matrix. Output the position component in the state posterior estimate as the filtered result of the vehicle's latitude and longitude coordinates: K k =P k|k-1 H T (HP k|k-1 H T +R k ) -1 ; P k|k =(IK k H)P k|k-1 .

[0055] The B-spline curve fitting algorithm is used to smooth the historical trajectory data of the vehicle after being processed by the Kalman filter algorithm to remove the mutations in the trajectory data. Specifically, the vehicle trajectory coordinate sequence processed by the Kalman filter is Treated as two-dimensional scattered data. Introducing parameter t i , parameterize the scattered data into where t i It can be i or the normalized chord length. According to the number of scattered points and the desired smoothness, the number of control points m of the B-spline curve is determined. According to the selected B-spline order (usually 3), the B-spline basis function matrix N∈R is calculated. n×m , n is the number of scattered points, m is the number of control points. Least squares solves the coordinates of the control points and constructs the least squares problem for the x and y coordinates respectively: NP x =X;NP y =Y, where P x and P y are the x and y coordinate vectors of the control point, and X and Y are the x and y coordinate vectors of the scattered data. Solve the above least squares problem to obtain the control point coordinates P x and P y Generate a B-spline curve, and generate the fitted B-spline curve C(t)=[NP x ,NP y ] T The vehicle trajectory data is smoothed using the B-spline curve fitting algorithm to remove mutations and jumps in the trajectory, thus obtaining a smooth and realistic driving trajectory, laying the foundation for subsequent feature extraction.

[0056] According to the battery model and capacity parameters, the SOC data of different batteries are normalized; for each battery model, the SOC data is mapped to the [0, 1] interval according to its rated capacity parameters. The normalization formula is:

[0057] Among them, SOC normalized is the normalized SOC value, SOC is the original SOC value, and SOC min and SOC max The SOC theoretical minimum and maximum values ​​of this battery model are respectively. The local anomaly factor LOF algorithm is used to detect anomalies in the normalized battery parameters; the specific implementation steps are as follows: Feature extraction, for each battery, extract its key parameters such as SOC, temperature, voltage, etc. to form a d-dimensional feature vector x i =[x i1 ,x i2 ,......,x id ] T. The eigenvectors of all batteries are combined into a eigenmatrix X = [x1, x2, ..., x n ] T , where n is the number of battery samples.

[0058] Calculate k-distance, for each feature vector x i , calculate the Euclidean distance between it and all other eigenvectors, and get the distance matrix D = [d ij ] n×n , where d ij =||x i -x j ||2. For each eigenvector x i , find the distance to its kth nearest neighbor, denoted as k-dist(x i ). Calculate the reachability distance, for each pair of feature vectors x i and x j , calculate the reach-dist distance between them: k (x i ,x j )=max(k-dist k (x i ),d ij ). Calculate the local reachability density, for each eigenvector x i , calculate its local reachability density: Where N k (x i ) represents x i The k nearest neighbor points of . Calculate the local anomaly factor, for each feature vector x i , calculate its local anomaly factor: Outlier determination, set the outlier determination threshold λ, for each feature vector x i , if LOF k (x i )>λ, then determine x i The battery samples that are judged as abnormal points are removed from the data set to obtain the battery parameter data after cleaning.

[0059] Figure 3This is an exemplary flow chart for generating vehicle history trajectory view features shown in the present application. The pre-processed vehicle history trajectory data and battery parameters are regarded as different views, and the corresponding features are extracted: According to the GPS positioning data in the vehicle history trajectory data, the distance between adjacent trajectory points is calculated to obtain the mileage of each trajectory segment, and a mileage sequence is generated. For each vehicle's GPS trajectory point sequence, two adjacent trajectory points form a trajectory segment. Using the conversion formula from longitude and latitude coordinates to spherical distance, the spherical distance of each trajectory segment is calculated as the mileage of the trajectory segment. The mileage of each trajectory segment is composed of a mileage sequence {d1, d2, ..., d n Set sliding windows of different sizes, perform sliding aggregation on the mileage sequence, and accumulate and sum the mileage in each sliding window to obtain a multi-scale mileage feature sequence. Set the size of the sliding window to

[0060] {w1,w2,......,w m}, where w1<w2<......<w m For scale w i , sliding size w on the mileage sequence with a step size of 1 i The window is used to calculate the mileage accumulation in each window to form a scale w i The mileage feature sequence {s i (1),s i (2),......,s i (nw i +1)}. By setting sliding windows of different scales, the mileage variation patterns of vehicles in different time spans can be captured.

[0061] Using the statistical feature extraction algorithm, the mileage feature sequence at each scale is extracted to obtain the mileage statistical features at the corresponding scale. i The mileage characteristic sequence under

[0062] {s i (1),s i (2),......,s i (nw i +1)}, calculate its mean u i , variance v i and kurtosis k i :

[0063]

[0064] The scale w is composed of mean, variance and kurtosis i The mileage statistical feature vector {ui ,v i ,k i}. Traverse sliding windows of different scales to obtain a set of multi-scale mileage statistical features. For m different sliding window scales {w1,w2,......,w m}, respectively perform the feature extraction steps of S32 and S33. Finally, a set of multi-scale mileage statistical feature vectors is obtained

[0065] {{u1,v1,k1},{u2,v2,k2},......,{u m ,v m ,k m The feature cascade algorithm is used to cascade the multi-scale mileage statistical features in scale order to generate the vehicle history trajectory view features. The mileage statistical feature vectors extracted at different scales are spliced ​​in order from small to large scale to form a 3m-dimensional feature vector: {u1,v1,k1,u2,v2,k2,.....,u m ,v m ,k m}, the above feature vector is the extracted vehicle history trajectory view feature, denoted by X traj .

[0066] Figure 4 This is an exemplary flow chart of generating battery parameter view features shown in the present application. According to the preprocessed battery parameter data, the battery charge and discharge characteristics are extracted as the battery parameter view features, including: according to the preprocessed battery parameter data, the battery SOC data is divided into charge and discharge cycles using a cycle detection algorithm to obtain each complete charge and discharge cycle data. The specific implementation steps are as follows: Analyze the preprocessed battery SOC time series data {s1, s2, ..., s n}, where s i Represents the SOC value at the i-th time point. Detect the maximum and minimum points of the SOC sequence: For the i-th time point, compare its SOC value s i With adjacent time points s i-1 and i+1 If s i-1 <s i >s i+1 , then s i is a maximum point of SOC, denoted as Where j is the number of the maximum point. If s i-1 >s i <s i+1 , then s i is a minimum point of SOC, denoted by Where j is the number of the minimum point. Determine the charge and discharge status: Set the SOC change threshold δ charge and δ discharge , respectively represent the judgment thresholds of charging and discharging states. and (The extreme point can be a maximum point or a minimum point), calculate the SOC change between them If Δs j >δ charge , then it is believed that and The charging process occurs between Δs j <-δ discharge , then it is believed that and The discharge process occurs between the two. Extracting the charge and discharge cycle data: Take a maximum point To the next maximum point For a complete charge and discharge cycle. and The SOC subsequence data between i , represents the SOC data of the i-th charge and discharge cycle. Traverse all adjacent maximum point pairs to obtain the battery SOC charge and discharge cycle set {C1, C2, ..., C k}, where k is the total number of charge and discharge cycles.

[0067] SOC data resampling, for each charge and discharge cycle i (i = 1, 2, ..., N), the following operations are performed: Get the current cycle SOC sequence C i The length of i ; Calculate the resampling factor Initialize a zero sequence C of length L i ', used to store the resampled SOC sequence; for each position j (j = 0, 1, ..., L-1) of the resampled SOC sequence: calculate the position of the original SOC sequence corresponding to the position: p = j × s i ; If p is an integer, the pth element of the original SOC sequence is directly assigned to C i '[j]; If p is a decimal, C is obtained by interpolation in the following way i '[j] value: Let the integer part of p be k and the decimal part be f, then k and k+1 are the adjacent positions of p in the original SOC sequence; if linear interpolation is used, then

[0068] C i '[j]=(1-f)×C i [k]+f×C i[k+1]; if spline interpolation is used, C is calculated using the cubic spline interpolation function i '[j], the specific formula is omitted. The SOC sequence C after resampling of the i-th charge and discharge cycle is obtained i '; Output the resampled SOC sequence C i '(i=1,2,......,N).

[0069] SOC curve fitting and feature extraction, for the resampled SOC curve C of the i-th charge and discharge cycle i ', perform the following operations: construct a time variable sequence t, with a length of L, representing the normalized time, with values ​​ranging from 0 to 1: Use the least squares method to fit the n-order polynomial and get the polynomial coefficients a0, a1, ..., a n , the fitting formula is:

[0070] SOC=a0+a1×t+a2×t 2 +......+a n ×t n . Calculate SOC range:

[0071] SOC_range i =max(C i ')-min(C i '). Calculate the average SOC change rate: Get the start time of the i-th cycle and end time Calculate the average SOC change rate: Calculate the cumulative discharge amount: Define the cumulative discharge amount function: f(t) = a1+2a2t+......+na n t n-1 , calculate the definite integral Get the absolute value of the cumulative discharge capacity SOC_range i , the specific integration method can use numerical integration, such as trapezoidal method or Simpson method. Generate the charge and discharge characteristic vector F of the i-th cycle i :F i ={soc range_i ,soc rate_i ,soc discharge_i}, output the feature vector F of all charge and discharge cycles i(i=1,2,......,N). The polynomial fitting part solves the polynomial coefficients by the least squares method. The polynomial order n can be selected according to the complexity of the SOC curve. Usually, 3 to 6 can better fit the SOC curve. The feature extraction part first calculates the SOC range, that is, the maximum change amplitude of SOC within the cycle. Then, the average SOC change rate is calculated according to the SOC range and the time span of the cycle, reflecting the speed of SOC change. Finally, the SOC change rate function is obtained by deriving the fitting curve, and then integrating it to obtain the cumulative discharge amount, which reflects the total amount of battery discharge within the cycle. These three features characterize the changing characteristics of SOC within a charge and discharge cycle from different angles, and can better characterize the charge and discharge characteristics of the cycle. Combining them into a feature vector can be used for subsequent clustering analysis and battery health status assessment.

[0072] Extract the feature vectors of all charge and discharge cycles and arrange them in order to obtain the time series of battery charge and discharge characteristics {F1, F2, ....., F k The sliding window method is used to calculate the mean, variance and skewness of each feature at different time scales, and generate multi-scale battery charge and discharge statistical features as battery parameter view features. Set sliding windows of different sizes {w1,w2,.....,w m}, sliding on the charge and discharge feature time series. For feature j (j = 1, 2, 3) and window size w i , calculate the corresponding mean u ij , variance v ij and skewness s ij :

[0073]

[0074] Finally, we get the battery parameter view feature X bat , which contains the mean, variance, and skewness of each charge and discharge feature at all sliding window scales.

[0075] Figure 5This is an exemplary flow chart for obtaining vehicle-battery comprehensive features shown in this application. The multi-view subspace clustering algorithm is used to fuse the vehicle history trajectory view features and the battery parameter view features to obtain the vehicle-battery comprehensive features: the PCA algorithm is used to reduce the dimension of the original features of the vehicle history trajectory view. For each vehicle, its historical trajectory GPS sequence, speed sequence and other original feature data are combined into a feature matrix X1 with a shape of (n, d), where n is the number of trajectory data points and d is the dimension of the original feature. The feature matrix X1 is centralized: the mean of each feature is calculated to obtain the mean vector u with a shape of (d,). The mean vector u is subtracted from each row of X1 to obtain the centralized feature matrix X1. 1_centered . Calculate the centralized feature matrix X 1_centered The covariance matrix C of is: Perform eigenvalue decomposition on the covariance matrix C: calculate the eigenvalues ​​and corresponding eigenvectors of C, sort the eigenvalues ​​from large to small, obtain the sorted eigenvalue vector λ and the corresponding eigenvector matrix W, select the eigenvectors corresponding to the first k1 largest eigenvalues, and construct the PCA projection matrix P1: P1 = W[:,:k1], where W[:,:k1] means taking all rows (:) and the first k1 columns (:k1) of the matrix W. The centralized eigenmatrix X 1_centered Mapped to the k1-dimensional subspace defined by P1, the compressed trajectory view feature V1 is obtained: V1 = X 1_centered ×P1, output the compressed trajectory view feature V1, the shape is (n, k1). The ICA algorithm is used to reduce the dimension of the original features of the battery parameter view. For each vehicle, its battery capacity, voltage and other monitoring parameters are combined into a feature matrix X2, the shape is (n, d), where n is the number of monitoring data points and d is the dimension of the original feature. The feature matrix X2 is centralized: the mean of each feature is calculated to obtain the mean vector u, the shape is (d,) means that the mean vector u is a one-dimensional vector containing d elements, d represents the number of elements, can be a row vector or a column vector, and has no specific direction. Subtract the mean vector u from each row of X2 to obtain the centralized feature matrix X 2_centered . For the centralized feature matrix X 2_centered Perform whitening: Calculate X 2_centered The covariance matrix C of C is decomposed by eigenvalue to obtain the eigenvalue vector λ and the eigenvector matrix W, and the whitening matrix is ​​calculated. Where diag represents a diagonal matrix, X 2_centered Multiply by the whitening matrix P whiten , get the whitened feature matrix X 2_centered .

[0076] Use the FastICA algorithm to estimate X2_centered Independent components in: Randomly initialize a matrix W of shape (d, k2) ica . Iteratively update W ica , until convergence or the maximum number of iterations is reached: Calculate W ica Update volume:

[0077] Update W ica :

[0078] W ica '=W ica +α×ΔW (α is the learning rate); for W ica Each column of is normalized to obtain the estimated independent component matrix W ica , the whitening matrix P whiten and the independent component matrix W ica Multiply them together to get the ICA projection matrix P2:

[0079] P2=P whiten ×W ica . The centralized feature matrix X 2_centered Mapped to the k2-dimensional subspace defined by P2, the compressed battery view feature V2 is obtained: V2 = X 2_centered ×P2, output the compressed battery view feature V2, the shape is (n, k2). Calculate the cosine similarity between the compressed vehicle history trajectory view feature V1 and the battery parameter view feature V2, and construct the inter-view similarity matrix S, initialize a similarity matrix S with a shape of (n, n), which is used to store the cosine similarity of the compressed features of each pair of vehicles on the two views.

[0080] For each vehicle i (i=1, 2, ..., n), calculate its compressed feature vector V on V1 and V2 1i and V 2i Cosine similarity between: Extract the compressed feature vector of the i-th vehicle on V1 and V2: V 1i = V1[i,:] (shape is (k1,)), V 2i = V2[i,:] (shape is (k2,)). Calculate V 1i and V 2i The L2 norm (Euclidean norm) of: Calculate V 1i and V 2i The inner product (dot product): dot product =∑(V 1i ×V 2i ).

[0081] Calculate V1i and V 2i The cosine similarity between: The cosine similarity S i Assign values ​​to the diagonal elements of the similarity matrix S: S[i,i] = S i For each pair of different vehicles i and j (i, j = 1, 2, ..., n; i ≠ j), calculate their compressed feature vector V on V1 and V2 1i , V 2i , V 1j , V 2j Cosine similarity between: Extract the compressed feature vectors of the i-th and j-th vehicles on V1 and V2: V 1i = V1[i,:] (shape is (k1,)), V 2i =V2[i,:](shape is (k2,)), V 1j =V1[j,:](shape is (k1,)), V 2j = V2[j,:] (shape is (k2,)). Calculate V 1i , V 2i , V 1j , V 2j The L2 norm of :

[0082] Calculate V 1i , V 2i , V 1j , V 2j The inner product of: dot product_ij =∑(V 1i ×V 2j ), dot product_ji =∑(V 1j ×V 2i ). Calculate V 1i , V 2i , V 1j , V 2j The cosine similarity between: The cosine similarity S ji and S ij Assign values ​​to the corresponding elements of the similarity matrix S: S[i,j] = S ij , S[j,i]=S ji , output the similarity matrix S between views, with shape (n, n).

[0083] Construct the multi-view reconstruction error objective function, objective function: Among them: X1 and X2 are the original feature matrices of the vehicle history trajectory view and the battery parameter view respectively; V1 and V2 are the compressed vehicle history trajectory view features and battery parameter view features respectively; P1 and P2 are the projection matrices corresponding to V1 and V2 respectively; a and b are the reconstruction error weights of the two views; ||·||_F represents the Frobenius norm of the matrix. The purpose of this objective function is to minimize the error of the compressed features V1 and V2 to reconstruct the original features X1 and X2. By minimizing the reconstruction error, we can learn a compressed feature representation that can better represent the original data. The inter-view similarity matrix S is introduced, the regularization term is constructed, and the overall optimization objective function of the multi-view subspace representation learning is generated. The overall optimization objective function is:

[0084] Where: V is the n×(k1+k2)-dimensional joint compressed feature matrix formed by concatenating V1 and V2 in rows; D is the diagonal matrix of S, that is tr(*) represents the trace of the matrix; c is the weight coefficient of the regularization term. Based on S34, S35 introduces the inter-view similarity matrix S and constructs a regularization term: c×tr(V T (DS)V). The purpose of this regularization term is to minimize the graph Laplacian operator of the compressed features, so that the samples with high similarity between views have closer learned compressed features. Specifically, the graph Laplacian operator L = DS, where D is the diagonal matrix of S. tr(V T (DS)V) can be expanded as:

[0085]

[0086] The alternating optimization strategy is used to solve the overall optimization objective function, randomly initialize the compression feature matrices V1 and V2, and randomly initialize the projection matrices P1 and P2. Repeat the following steps until the overall objective function converges or the maximum number of iterations is reached: Fix P1 and P2, optimize V1 and V2: Concatenate V1 and V2 by row to form a joint compression feature matrix V, and calculate the diagonal matrix D of S, where Construct the optimization subproblem of V: Solve for the closed-form solution of V: Extract the optimized V1 and V2 from V. Fix V1 and V2, optimize P1 and P2: Construct the optimization sub-problem of P1 (ridge regression problem): Solve for the closed-form solution of P1: Construct the optimization subproblem of P2 (ridge regression problem): Solve for the closed-form solution of P2: Calculate the current overall objective function value and determine whether it has converged. Output: optimized compressed feature matrices V1 and V2, optimized projection matrices P1 and P2.

[0087] Specifically, we alternately fix the compressed feature matrix V and the projection matrix P, and optimize the other parameter each until the objective function converges. When P is fixed, the subproblem of optimizing V is a quadratic function. We can get the closed-form solution of V by taking the derivative, which is: This closed-form solution can directly calculate the optimized V without iterative optimization. When V is fixed, the subproblem of optimizing P becomes a classic ridge regression problem. For P1 and P2, we construct the corresponding ridge regression problems and solve their closed-form solutions: Among them, λ is the regularization parameter of ridge regression, which is used to control the complexity of the model and prevent overfitting. On each vehicle, we iteratively perform the optimization of the above two sub-problems until the overall objective function converges. In this way, we obtain the optimized compressed features V1 and V2 of the current vehicle. Finally, the above optimization process is performed on all vehicles to obtain the optimized compressed features V1 and V2 of all vehicles. These optimized features are concatenated row by row to form a comprehensive vehicle-battery feature representation. This alternating optimization strategy takes advantage of the convexity and decomposability of the objective function. By decomposing the original problem into multiple sub-problems and alternately optimizing different sets of parameters, the optimization process is effectively simplified. At the same time, since each sub-problem has an explicit closed-form solution, complex iterative optimization is avoided and the computational efficiency of the algorithm is improved.

[0088] Figure 6 This is an exemplary flow chart for obtaining clustering results of vehicle-battery comprehensive features shown in this application, including: calculating the vehicle-battery similarity matrix, input: vehicle-battery comprehensive feature vectors obtained by clustering, there are n vehicle-battery pairs, the dimension of each feature vector is m, and an n×n vehicle-battery similarity matrix D is initialized to store the similarities between different vehicle-battery comprehensive features. For each pair of vehicle-battery comprehensive feature vectors (i, j), i, j = 1, 2, ..., n: calculate the Pearson correlation coefficient r between the i-th vehicle-battery comprehensive feature vector and the j-th vehicle-battery comprehensive feature vector ij , r ij As the element D in the vehicle-battery similarity matrix D ij and D ji , Output: vehicle-battery similarity matrix D. Generate hierarchical clustering tree, initialize cluster set, input: vehicle-battery similarity matrix D, size is n×n, initialize cluster set C0 = {{1}, {2}, ....., {n}}, where {i} represents the cluster containing only the i-th vehicle-battery pair, output: initial cluster set C0. Merge the two most recent clusters, input: current cluster set C t , vehicle-battery similarity matrix D. For C tFor any two clusters A and B in the same cluster, calculate their single connection distance: d(A,B) = min{i∈A,j∈B}D ij , in C t Find the two clusters A and B with the smallest single connection distance, merge A and B into a new cluster M = A*∪B*, and update the cluster set C t+1 =C t \{A*,B*}∪{M}, which means merging A* and B* into a new cluster M, and then removing A* and B* from the cluster set C t Finally, add the new cluster M to the updated cluster set C. t+1 Output: The merged cluster set C t+1 . Update the similarity matrix, input: the merged cluster set C t+1 , vehicle-battery similarity matrix D. For the newly merged cluster M, calculate its similarity with C t+1 The single connection distance of other clusters in the cluster is used to update the similarity matrix D: Output: Updated vehicle-battery similarity matrix D. Repeat the merging of clusters until a complete hierarchy is formed, and repeat until only one element remains in the cluster set, namely C T ={1,2,......,n}, output: a series of nested cluster sets {C0,C1,......,C T}.

[0089] Construct a hierarchical clustering tree. Input: a series of nested cluster sets {C0, C1, ..., C T}, each independent cluster {i} in the initial cluster C0 is used as a leaf node of the hierarchical clustering tree. For each new cluster M generated by cluster merging: M is added as an internal node to the hierarchical clustering tree, the two merged clusters A and B are used as child nodes of M, and the final cluster containing all vehicle-battery pairs is used as the root node of the hierarchical clustering tree. Output: a hierarchical clustering tree T consisting of leaf nodes, internal nodes and root nodes. A genetic algorithm is used to search for the global optimal partition of the hierarchical clustering tree, and a global fitness function is defined. Input: hierarchical clustering tree T, vehicle-battery similarity matrix D, for a partitioning scheme P of the hierarchical clustering tree T, k clusters {C1, C2, ..., C k}. Define the intra-cluster similarity S in : That is, for each cluster C i , calculate the sum of similarities between all vehicle-battery pairs within a cluster, and then sum over all clusters. Define the inter-cluster similarity S out : That is, for each pair of different clusters C i and C j, calculate the sum of similarities between all vehicle-battery pairs, and then sum all cluster pairs. Define the global fitness function f(P): That is, the ratio of the similarity within a cluster to the similarity between clusters. Output: global fitness function f(P).

[0090] Generate the initial partition population. Input: hierarchical clustering tree T, initial population size n, number of partition points m. Repeat the following steps n times to generate the initial population P0: randomly select m internal nodes in the hierarchical clustering tree T as partition points, treat these m nodes and their subtrees as independent clusters, and the rest of T as another cluster to form a partition individual. Output: initial population P0 containing n partition individuals. Calculate individual fitness. Input: initial population P0, global fitness function f(P). For each partition individual P in population P0, i :Calculate its fitness value f(P according to the defined global fitness function i ), output: the fitness value corresponding to each partitioned individual. Genetic algorithm iterative optimization, input: current population P t , maximum number of iterations max_iter, fitness threshold threshold, repeat the following steps until the maximum number of iterations is reached or the average fitness of the population reaches the threshold: Selection: According to the individual fitness value, use methods such as roulette or tournament selection to select from P t Select a certain number of excellent individuals from P, crossover: perform crossover operation on the selected excellent individuals, randomly exchange some of their partitioning points, and generate new partitioning individuals, mutation: mutate the newly generated partitioning individuals with a certain probability, randomly change some of the partitioning points, update: replace P with the newly generated partitioning individuals. t The individuals with lower fitness in the population are obtained as the new population P t+1 , output: the final partition population P T .

[0091] Select the optimal partition result, input: the final partition population P T , the global fitness function f(P), in P T Select the partition individual P* with the largest global fitness function value, and cluster the corresponding cluster results {C1, C2*, ....., C k *} as the global optimal division of the hierarchical clustering tree T, that is, the final clustering result of the vehicle-battery comprehensive characteristics. Output: Clustering result of vehicle-battery comprehensive characteristics {C1,C2*,.....,C k *}.

[0092] Figure 7is an exemplary flow chart of constructing a battery remaining capacity prediction model shown in the present application, and the battery remaining capacity prediction model is trained using a decision tree algorithm, specifically including: constructing a training sample set, for each cluster C i (i=1, 2, ..., k), perform the following steps: extract the historical trajectory data and battery parameter data of all vehicles in the cluster. The historical trajectory data includes: vehicle ID, timestamp, GPS location, speed, acceleration, fuel consumption, etc. The battery parameter data includes: vehicle ID, timestamp, voltage, current, temperature, remaining capacity, etc. For cluster C i For each vehicle in the range, the following processing is performed: the historical trajectory data and battery parameter data of the vehicle are aligned according to the timestamp to form time series data. For each timestamp t j (j=1, 2, ..., n_i), extract the following features and labels: feature x ij :Vehicle historical trajectory features: Extract timestamp t j The trajectory data in the previous window is used to calculate the average speed, average acceleration, total mileage, average fuel consumption and other features. Battery parameter features: extract timestamp t j Battery parameter values ​​such as voltage, current, and temperature at the moment. ij : Extract timestamp t j The remaining capacity value of the battery at the moment, the feature-label pair (x ij ,y ij ) is stored as a set of training samples. Cluster C i The training samples of all vehicles in the cluster are summarized to form the training sample set D of the cluster. i :{D i =(x ij ,y ij ),j=1,2,.....,n i}, where n i Represents cluster C i The total number of samples of all vehicles in the cluster. Preprocess the training sample set D_i: normalize or standardize the features, scale the feature values ​​to a similar range, perform necessary conversion or normalization on the labels, and divide the training set and validation set into a certain proportion for model training and performance evaluation. Output: The training sample set corresponding to each cluster {D1,D2,.....,D k}, each training sample set contains multiple sets of feature-label pairs (x ij ,y ij ).

[0093] The CART decision tree algorithm is used to train the battery remaining capacity prediction model and build a CART decision tree model. Input: training sample set D i, feature set A, target variable y (remaining battery capacity), initialize the root node of the decision tree, and train the sample set D i Input to the root node. Recursively perform feature selection and sample partitioning on each node to generate a decision tree: For the current node t, select the optimal split feature A according to the Gini index criterion k *, according to A k *The value of the current node sample set D t Divided into sample sets D1 and D2 of left and right child nodes r , the left and right child nodes are recursively split until one of the decision tree growth stop conditions is met. Output: initial CART decision tree model. The Gini index is used as the evaluation criterion. Input: current node t, node sample set D_t, feature set A. For each feature A k (k=1,2,......,m), calculate its Gini index: According to A k The value of D t Divide into V subsets For each subset Calculate its Gini impurity: Among them, C is the number of categories of the target variable y, For subset The number of samples where the target variable takes the value c. Calculate feature A k The Gini index is: Select feature A with the smallest Gini index k *As the optimal splitting feature, according to A k *The value of D t Divided into sample sets D1 and D2 of left and right child nodes r Output: optimal split feature A k *, left and right child node sample sets D1 and D r .

[0094] Recursively split the node until the stop condition is met, input: current node t, node sample set D t , decision tree growth stop condition. Determine whether one of the decision tree growth stop conditions is met: the number of samples at the current node | D t | is less than the preset sample threshold, the samples of the current node have the same value on the target variable y, the feature values ​​of the current node are the same, the depth of the decision tree reaches the preset maximum depth, the minimum Gini index of the current node is greater than or equal to the Gini index of the parent node, if the stopping condition is met, mark the current node as a leaf node, if the stopping condition is not met, recursively split the left and right child nodes respectively, output: the generated decision tree subtree. Mark the leaf node and calculate the output value, input: the node t that meets the decision tree growth stopping condition, corresponding to the sample set D t, mark node t as a leaf node. Calculate the output value of leaf node t: Among them, |D t | is the number of leaf node samples, y j is the target variable value of the sample (remaining battery capacity). Output: Node t marked as a leaf node, output value

[0095] To prune leaf nodes, input: initial CART decision tree model, validation set, pruning threshold, use cross-validation method, and use the validation set to evaluate the pruning effect of each leaf node. For each leaf node t, calculate the change in loss function before and after pruning: Among them, R(t) is the error sum of leaf node t, R(T t ) is the subtree T with t as the root t The sum of the squared errors of the leaf nodes, |T t | is the number of leaf nodes in the subtree. For leaf nodes whose ΔL(t) is less than the preset threshold: merge it with the corresponding parent node into a new leaf node, update the decision tree structure, and repeat the pruning process until all ΔL(t) are greater than the threshold. Output: the final CART decision tree model. Use the cross-validation method to optimize the decision tree hyperparameters. S531: Use grid search to determine the candidate value range of the hyperparameters, and determine the decision tree hyperparameters: maximum depth and minimum number of samples for leaf nodes. Set the candidate value range for the maximum depth: minimum value d min , maximum value d max , interval Δd, generate a candidate value sequence: {d min ,d min +Δd,d min +2Δd,......,d max}, set the candidate value range for the minimum number of samples for leaf nodes: minimum value n min , the maximum value n max , interval Δn. Generate candidate value sequence: {n min ,n min +Δn,n min +2Δn,......,n max Generate a hyperparameter grid: Combine the candidate value sequences of maximum depth and minimum number of samples to obtain a grid of hyperparameter candidate value combinations (d, n). Output: a hyperparameter grid, where each grid point corresponds to a set of hyperparameter candidate value combinations (d, n).

[0096] Use K-fold cross validation method to divide the training set and validation set. Input: training sample set D i , fold number K, the training sample set D i Randomly divide into K mutually exclusive subsets: Perform K cross validations, and in each iteration: select K-1 subsets to form the training set D train , the remaining subset is used as the validation set D val Output: K-fold cross-validation training set and validation set partitioning results. For each set of hyperparameter candidate value combinations, train and evaluate the CART decision tree model. Input: hyperparameter grid, K-fold cross-validation training set and validation set partitioning results. For each set of hyperparameter candidate value combinations (d, n): D train is the training set, and the CART decision tree model M(d,n) is trained according to the validation set D val Evaluate the prediction performance of the model M(d,n) by calculating the root mean square error (RMSE): in, For the model to sample x j The prediction output records the cross-validation performance RMSE under the current hyperparameter combination (d, n) k (d,n). Output: K-fold cross validation performance {RMSE k (d,n),k=1,2,.....,K}.

[0097] Select the hyperparameter combination with the best model performance to obtain the optimized CART decision tree model. Input: hyperparameter grid, K-fold cross-validation performance corresponding to each set of hyperparameter candidate value combination (d, n), and for each set of hyperparameter candidate value combination (d, n): calculate the average performance of K-fold cross-validation: Select the hyperparameter combination (d*,n*) with the best average performance: Take (d*,n*) as the hyperparameter and use the entire training sample set D i The optimized CART decision tree model M* is obtained through training. Output: The optimized CART decision tree model M* is used as the battery remaining capacity prediction model in the cluster. The battery remaining capacity of the vehicles in the cluster is predicted using the optimized battery remaining capacity prediction model.

[0098] According to the clustering results of vehicles and batteries and the prediction results of the remaining battery capacity in the corresponding cluster, the battery recycling strategy for each vehicle-battery cluster is generated, and the battery recycling time window corresponding to each cluster is determined. For each vehicle-battery cluster C i :Get the predicted remaining capacity of all vehicle batteries in the cluster in the future time period: For cluster C i For each vehicle battery j (j=1, 2, ..., n_i), use the trained battery remaining capacity prediction model to predict its remaining capacity Q in the future time period ij (t), get cluster C iThe predicted remaining capacity of all vehicle batteries in the ij (t),j=1,2,......,n i}. Calculate cluster C based on the predicted remaining capacity i Distribution of battery remaining capacity in each time period: For each future time period t, statistical cluster C i The distribution of the remaining battery capacity in the battery is calculated, and the frequency or probability of the remaining battery capacity in different intervals is calculated to obtain the remaining capacity distribution P i (Q,t),P i (Q,t) means at time t, cluster C i The probability or frequency that the remaining capacity of the battery is Q. Set one or more battery remaining capacity thresholds: According to the needs of battery recycling, set one or more battery remaining capacity thresholds {Q1, Q2, ..., Q m}, threshold Q k Indicates the remaining capacity status of the battery when it needs to be recycled. For example, it needs to be recycled when the remaining capacity drops below 30%.

[0099] Calculate cluster C i The time required for the remaining battery capacity to decrease to each threshold: For each threshold Q k (k=1, 2, ..., m), according to cluster C i The remaining battery capacity distribution P i (Q,t), it is estimated that the remaining capacity of the battery in the cluster is reduced to Q k The time required t ik , can be obtained by interpolation, fitting and other methods according to P i The changing trend of (Q,t) can be used to estimate t ik , we get cluster C i The time required for the remaining battery capacity to drop to each threshold is: ik ,k=1,2,......,m}. Determine the cluster C i Battery recycling time window: According to {t ik ,k=1,2,......,m}, determine the cluster C i Battery recycling time window Indicates the earliest time when the batteries in the cluster need to be recycled, which can be taken as t ik The minimum value among or other suitable values, Indicates the latest time that the batteries in the cluster need to be recycled, which can be taken as t ik The maximum value or other suitable value in the battery recycling time window Represents cluster C i The vehicle batteries need to be recycled within this time frame, and each cluster C is output iCorresponding battery recycling time window Plan and deploy differentiated battery recycling outlets to obtain the historical driving trajectory distribution of each vehicle: For each vehicle v j (j=1, 2, ..., n), obtain its historical driving GPS trajectory data {(x jt ,y jt ),t=1,2,......,T j},(x jt ,y jt ) represents vehicle v j At the location coordinate at time t, by analyzing the historical trajectory data, the driving trajectory distribution of each vehicle is obtained, reflecting the vehicle's activity area and movement mode. Calculate the spatial clustering area of ​​each vehicle-battery cluster: For each vehicle-battery cluster C i (i=1, 2, ..., k), calculate the spatial clustering area A of the vehicle trajectory distribution within the cluster i , A i Represents cluster C i The main area of ​​vehicle activity can be extracted from the vehicle trajectory distribution by clustering, density estimation and other methods to obtain the set of spatial aggregation areas of each cluster {A i ,i=1,2,......,k}.

[0100] Planning and deploying battery recycling outlets: Comprehensively consider the spatial aggregation areas of each cluster {A i ,i=1,2,......,k} and the spatial distribution of vehicle trajectories in the entire domain. In areas where vehicle trajectories are dense and multiple clustering spatial aggregation areas overlap, key recycling outlets are deployed. In these areas, vehicle activities are frequent and the distribution of vehicle trajectories of multiple clusters overlaps. It is suitable to set up large or multiple recycling outlets. Key recycling outlets can serve multiple clustered vehicles to improve recycling efficiency and coverage. In areas where vehicle trajectories are relatively scattered and different clustering spatial aggregation areas are independent, decentralized recycling outlets are deployed. In these areas, vehicle activities are relatively independent and the distribution of vehicle trajectories of different clusters is relatively scattered. It is suitable to set up small or single recycling outlets. Decentralized recycling outlets can provide customized recycling services for specific types of vehicles. The specific location and number of recycling outlets can be determined through optimization algorithms, geographic information analysis and other methods, taking into account service coverage and operating costs. Develop differentiated battery recycling plans: For each cluster C i , formulate differentiated battery recycling plans according to their characteristics, battery recycling time: according to cluster C i Battery recycling time window Choose the appropriate recycling time. Within the scope, comprehensively consider factors such as vehicle usage and battery health status to determine the specific recycling time and recycling location: select cluster Ci The locations with high matching degree between vehicle trajectory distribution and recycling network points are selected as recycling locations. i The spatial relationship between vehicle trajectory distribution and recycling points can be used to select convenient and efficient recycling locations. The selection of recycling locations can reduce the transportation distance and cost of vehicle batteries and improve recycling efficiency.

Claims

1. A new energy vehicle battery recycling management method, comprising: Collect multi-source heterogeneous data of the vehicle, including vehicle historical trajectory data and battery parameters; battery parameters include SOC data; vehicle historical trajectory data includes GPS positioning data; Preprocess the collected multi-source heterogeneous data; The preprocessed vehicle history trajectory and battery parameters are regarded as different views, and the corresponding features are extracted: the vehicle's driving trajectory features are extracted as the vehicle history trajectory view features; the battery's charging and discharging features are extracted as the battery parameter view features; the vehicle history trajectory view features and battery parameter view features are fused using a multi-view subspace clustering algorithm to obtain the vehicle-battery comprehensive features; Based on the comprehensive characteristics of vehicle-battery, the similarity between the comprehensive characteristics of different vehicle-batteries is calculated, and a hierarchical clustering tree of vehicle-battery is constructed; according to the hierarchical clustering tree, the clustering result of vehicle-battery is obtained; According to the clustering results of vehicles and batteries, in each cluster, the vehicle historical trajectory data and battery parameter data are used as training samples, and the decision tree algorithm is used to train the battery remaining capacity prediction model; the trained battery remaining capacity prediction model is used to predict the battery remaining capacity of the vehicles in the cluster; Based on the clustering results of vehicles and batteries and the prediction results of the remaining capacity of the batteries in the corresponding clusters, a battery recycling strategy for each vehicle-battery cluster is generated.

2. The new energy vehicle battery recycling management method according to claim 1 is characterized in that: Preprocess the collected multi-source heterogeneous data, including: The Kalman filter algorithm is used to denoise the GPS positioning data in the vehicle historical trajectory data to eliminate GPS signal errors and drifts; The B-spline curve fitting algorithm is used to smooth the vehicle historical trajectory data processed by the Kalman filter algorithm to remove the sudden changes in the trajectory data; Normalize the SOC data of different batteries according to battery model and capacity parameters; The local outlier factor (LOF) algorithm is used to detect anomalies in the normalized battery parameters.

3. The new energy vehicle battery recycling management method according to claim 2 is characterized in that: Extract the vehicle's driving trajectory features as the vehicle's historical trajectory view features, including: According to the GPS positioning data in the historical trajectory data of the vehicle, the distance between adjacent trajectory points is calculated to obtain the mileage of each trajectory and generate a mileage sequence; Set sliding windows of different sizes, perform sliding aggregation on the mileage sequence, and accumulate and sum the mileage in each sliding window to obtain a multi-scale mileage feature sequence; Using statistical feature extraction algorithm, the mileage feature sequence at each scale is extracted to obtain the mileage statistical features at the corresponding scale. Traverse sliding windows of different scales to obtain a set of multi-scale mileage statistical features; The feature cascade algorithm is used to cascade the multi-scale mileage statistical features in scale order to generate the vehicle history trajectory view features.

4. The new energy vehicle battery recycling management method according to claim 3 is characterized in that: Extract the battery charging and discharging characteristics as battery parameter view features, including: According to the pre-processed battery parameter data, the battery SOC data is divided into charge and discharge cycles using a cycle detection algorithm to obtain data for each complete charge and discharge cycle; For each charge and discharge cycle data, the SOC data is resampled using an interpolation algorithm to obtain an SOC curve with consistent length; The SOC curve is fitted using a polynomial function, and the SOC range, SOC change rate and cumulative discharge capacity are extracted as a set of characteristic vectors corresponding to the cycle charge and discharge characteristics; Extract the feature vectors of all charge and discharge cycles and arrange them in order to obtain the time series of battery charge and discharge characteristics; The sliding window method is used to calculate the mean, variance and skewness of each feature at different time scales to generate multi-scale battery charging and discharging statistical features as battery parameter view features.

5. The new energy vehicle battery recycling management method according to claim 4 is characterized in that: The multi-view subspace clustering algorithm is used to fuse the vehicle history trajectory view features and the battery parameter view features to obtain the comprehensive features of the vehicle-battery, including: The principal component analysis algorithm PCA is used to reduce the dimension of the vehicle history trajectory view features to obtain the compressed vehicle history trajectory view features; The independent component analysis algorithm ICA is used to reduce the dimension of the battery parameter view features to obtain the compressed battery parameter view features; Calculate the cosine similarity between the compressed vehicle history trajectory view features and the battery parameter view features to obtain the inter-view similarity matrix, which is used to reflect the correlation between different view features; Construct a multi-view reconstruction error objective function, and use the compressed vehicle history trajectory view features and battery parameter view features to reconstruct the original view features respectively; Based on the multi-view reconstruction error objective function, the inter-view similarity matrix is ​​introduced to construct a regularization term and generate the overall optimization objective function of multi-view subspace representation learning. The alternating optimization strategy is used to solve the overall optimization objective function and obtain the comprehensive characteristics of the vehicle-battery.

6. The new energy vehicle battery recycling management method according to claim 5 is characterized in that: Based on the comprehensive features of vehicle-battery, the similarity between the comprehensive features of different vehicle-batteries is calculated, and a hierarchical clustering tree of vehicle-battery is constructed; According to the hierarchical clustering tree, the clustering results of vehicle-battery are obtained, including: Based on the vehicle-battery comprehensive features obtained by clustering, the Pearson correlation coefficient is used to calculate the similarity between different vehicle-battery comprehensive features to obtain the vehicle-battery similarity matrix; According to the vehicle-battery similarity matrix, a single-link minimum distance algorithm is used to construct a vehicle-battery comprehensive feature hierarchical clustering tree; The global fitness function is set, and the genetic algorithm is used to search for the global optimal partition of the hierarchical clustering tree to obtain the clustering results of the vehicle-battery comprehensive characteristics.

7. The new energy vehicle battery recycling management method according to claim 6 is characterized by: Genetic algorithm is used to search for the global optimal partition of the hierarchical clustering tree, including: The global fitness function is set as the ratio of the sum of the similarity values ​​of vehicle-battery comprehensive characteristics within a cluster to the sum of the similarity values ​​of vehicle-battery comprehensive characteristics between clusters; Randomly select the partition points of the vehicle-battery comprehensive feature hierarchical clustering tree to generate the initial partition population, and each partition individual corresponds to a clustering result; For each individual in the initial partition population, calculate the corresponding global fitness function value as the individual fitness; The genetic algorithm is used to iterate the current population until the maximum number of iterations is reached or the average fitness of the population reaches a preset threshold; From the final partition population, the partition individual with the highest global fitness is selected, and the corresponding clustering result is used as the global optimal partition result of the hierarchical clustering tree to obtain the clustering result of the vehicle-battery comprehensive characteristics.

8. The new energy vehicle battery recycling management method according to claim 6 is characterized by: In each cluster, a decision tree algorithm is used to train a battery remaining capacity prediction model; including: In each cluster, the historical trajectory data of the vehicle and the battery parameter data at the corresponding time are extracted to construct a training sample set; The CART decision tree algorithm is used to train the battery remaining capacity prediction model. The training sample set is used as the model input, and the battery remaining capacity is used as the target variable. The decision tree is recursively generated through feature selection and sample partitioning. The cross-validation method is used to optimize the maximum depth and minimum number of node samples of the CART decision tree to obtain the optimized battery remaining capacity prediction model; The optimized battery remaining capacity prediction model is used to predict the battery remaining capacity of vehicles in the cluster.

9. The new energy vehicle battery recycling management method according to claim 8 is characterized by: The CART decision tree algorithm is used to train the battery remaining capacity prediction model, including: Construct a CART decision tree model, use the training sample set as the model input, take the remaining battery capacity as the target variable, and recursively generate a decision tree through feature selection and sample partitioning; The Gini index is used as the evaluation criterion to calculate the impurity of each feature in the sample division, and the feature with the smallest Gini index is selected as the optimal split feature. The current node sample set is split into a binary tree according to the value of the optimal split feature to form left and right child nodes. Recursively split the left and right child nodes until one of the decision tree growth stopping conditions is met; the decision tree growth stopping conditions include: the number of samples of the current node is less than the preset sample threshold; the samples of the current node have the same value on the target variable; the feature values ​​of the current node are the same; the depth of the decision tree reaches the preset maximum depth; the minimum Gini index of the current node is greater than or equal to the Gini index of the parent node; The nodes that meet the decision tree growth stop condition are marked as leaf nodes, and the average remaining battery capacity of the corresponding samples is used as the output value of the corresponding leaf node; The leaf nodes are pruned, and the loss function before and after pruning is calculated for each leaf node from bottom to top. When the reduction value of the loss function after pruning is less than the threshold, the corresponding leaf node and the corresponding parent node are pruned and merged into a new leaf node until the reduction value of the loss function of all leaf nodes after pruning is greater than the threshold, and the initial CART decision tree model is obtained.

10. The new energy vehicle battery recycling management method according to claim 9 is characterized in that: The cross-validation method is used to optimize the maximum depth of the decision tree and the minimum number of node samples to obtain the optimized battery remaining capacity prediction model, including: Use grid search to determine the candidate value range of hyperparameters, which include the maximum depth of the CART decision tree and the minimum number of samples for leaf nodes; The training sample set is divided into K mutually exclusive subsets, and K-1 subsets are selected as training sets each time, and the remaining subset is used as the validation set; For each set of candidate hyperparameter values, train and evaluate the performance of the CART decision tree model under the current parameter combination on the validation set; All candidate hyperparameter value combinations are traversed, and the hyperparameter combination with the best model performance is selected to obtain the optimized CART decision tree model as the battery remaining capacity prediction model.

Images