Adaptive electric vehicle battery state prediction and intelligent charging method and system based on graph neural network

Abstract

The application belongs to the technical field of electric vehicle battery management system, and particularly relates to a self-adaptive electric vehicle battery state prediction and intelligent charging method and system based on a graph neural network, which comprises the following steps: S1, battery monomer relationship graph construction; S2, SOC / SOH joint prediction based on a graph attention network; S3, self-adaptive charging strategy generation; and S4, self-adaptive learning mechanism. The application can improve prediction accuracy, optimize charging efficiency, improve battery consistency, enhance system robustness, has outstanding practical value, and has a good market application prospect.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of battery management (BMS) systems for electric vehicles, specifically involving battery state prediction based on deep learning, graph neural network modeling, and intelligent charging control technology. BACKGROUND

[0002] SOC (State of Charge): The ratio of the current remaining capacity to the rated capacity of the battery, usually expressed in percentage.

[0003] SOH (State of Health): The ratio of the current maximum available capacity to the initial rated capacity of the battery, reflecting the degree of battery aging.

[0004] Graph Neural Network (GNN): A deep learning model for processing graph-structured data, learning the representation of nodes and edges in the graph through message passing mechanisms between nodes.

[0005] Battery Cell Relationship Graph: A graph structure model constructed by taking each cell in the battery pack as a node, and physical connection relationships, thermal coupling relationships, and electrochemical correlations as edges.

[0006] Multi-Head Attention Mechanism: Through multiple attention heads, different dimensions of attention weights are calculated in parallel to enhance the model's ability to focus on key information.

[0007] Adaptive Charging Strategy: An intelligent control method that dynamically adjusts the charging parameters of each cell based on real-time state differences.

[0008] Traditional battery state prediction methods include: ampere-hour integration method, open-circuit voltage method, and equivalent circuit model. The ampere-hour integration method calculates SOC by integrating the charge and discharge current, but has cumulative error problems and the accuracy decreases over time. The open-circuit voltage method is based on the OCV-SOC relationship curve prediction, which requires a long resting time and is not suitable for dynamic conditions. The equivalent circuit model is to establish an RC equivalent circuit combined with Kalman filter prediction, which is complex in parameter identification and sensitive to battery types.

[0009] Machine learning methods include: neural network method, support vector machine, and ensemble learning. The neural network method uses MLP, RNN, LSTM, etc. models to predict SOC / SOH based on voltage, current, temperature, etc. The support vector machine solves the non-linear prediction problem through kernel function mapping. Ensemble learning combines multiple algorithms to improve prediction accuracy.

[0010] Existing battery state prediction methods mainly focus on a single battery or treat the battery pack as a whole, ignoring the mutual influence between cells in the battery pack, resulting in limited prediction accuracy (usually 3-5% error) and difficulty in supporting fine charging control.

[0011] Battery charging techniques include constant current constant voltage (CC-CV) charging, pulse charging technology and multi-stage charging. Constant current constant voltage (CC-CV) charging is charged at a fixed current to the cutoff voltage in the constant current stage, and the cutoff voltage is maintained until the current decays to the set value in the constant voltage stage. There is a problem of charging strategy solidification, which cannot adapt to the difference between single bodies. The pulse charging technology is superimposed on the basis of constant current with pulse signal, which can reduce the polarization effect to a certain extent, but the pulse parameter setting mainly relies on experience and lacks theoretical guidance. Multi-stage charging is charged with decreasing current steps, which can reduce the charging time compared with CC-CV, but the stage parameters are fixed, and the individualization degree is insufficient.

[0012] The existing charging method adopts a unified strategy, which cannot be personalized according to the state difference of single bodies, resulting in overcharging or insufficient charging of some single bodies, affecting the overall performance and life of the battery pack.

[0013] The application of graph neural network in battery management field is just starting, mainly focusing on battery thermal runaway early warning, fault diagnosis and other aspects, and the application in SOC / SOH prediction is less, and most of them are theoretical research, lacking of graph modeling method for battery pack structure characteristics, not fully utilizing the multi-dimensional correlation between single bodies, and lacking of method for converting graph neural network prediction results into charging strategy.

[0014] The information disclosed in this BACKGROUND section is only intended to increase an understanding of the general context in which the present application can be practiced. It is not admitted that any of the information provided in this section constitutes prior art against the present application. SUMMARY

[0015] The purpose of the present application is to provide a graph neural network-based adaptive electric vehicle battery state prediction and intelligent charging method and system to solve the problems of insufficient battery state prediction accuracy, difficulty in supporting fine charging control, ignoring the mutual influence relationship between single bodies in the battery pack, charging strategy solidification, inability to adapt to the state difference of single bodies, and lack of effective single body level charging control method in the prior art.

[0016] In order to achieve the above purpose, the present application provides the following technical solutions:

[0017] The graph neural network-based adaptive electric vehicle battery state prediction and intelligent charging method comprises the following steps:

[0018] S1, battery single body relationship graph construction, specifically comprising the following steps:

[0019] S11, for the single body in the battery pack, defining a node feature vector;

[0020] S12, define three types of edge relations, including: physical connection relationship, thermal coupling relationship and state correlation;

[0021] S13, multi-level adaptive weight fusion of node feature vector and three types of variable relations;

[0022] S2, SOC / SOH joint prediction based on graph attention network, specifically including the following steps:

[0023] S21, enhance multi-dimensional attention aggregation mechanism, adopt hierarchical-group attention calculation;

[0024] S22, intelligent adaptive feature aggregation;

[0025] S23, multi-objective intelligent joint prediction output;

[0026] S3, adaptive charging strategy generation, specifically including the following steps:

[0027] S31, intelligent adaptive SOC hierarchical strategy, dynamically determine hierarchical threshold according to GNN prediction result and battery characteristics;

[0028] S32, multi-factor intelligent current optimization;

[0029] S33, multi-level intelligent safety management and control, ensure that the charging current is within the dynamic safety range;

[0030] If the safety check passes, charging is performed; if it does not pass, the current is limited; the final current allocation is fed back to the GNN prediction input.

[0031] As preferred, in S11, for the i-th single cell in the battery pack, the node feature vector is defined as:

[0032] ;

[0033] Wherein, are voltage, current, and temperature measurement values; , are the change rates of voltage and temperature; is the internal resistance estimate value; is the state value at the previous time;

[0034] The physical connection relationship defined in S12 adopts an intelligent configurable weight model:

[0035] ;

[0036] Wherein, and respectively represent the connection function and the branch function, which are used to identify the physical connection relationship between single cells; The adaptive weight parameter vector is dynamically adjusted according to the battery pack topology structure; The intelligent connection function adopts a topology-aware weight calculation method:

[0037] ;

[0038] wherein, is a topology weight, is a branch weight, is a distance attenuation weight; calculation example: when =1, =0.8, the distance attenuation term =0.6, and =0.6, =0.3, =0.1, =0.6x1+0.3x0.8+0.1x0.6=0.84;

[0039] The thermal coupling relationship defined in S12 adopts a multi-scale adaptive diffusion model:

[0040] ;

[0041] wherein, is a multi-scale thermal coupling function, is a physical distance between monomers i and j, is a multi-scale adaptive diffusion parameter vector, is a temperature field distribution vector; A multi-level diffusion model is adopted:

[0042] ;

[0043] wherein, K is the number of diffusion scales, is a weight of each scale, is a diffusion parameter; calculation example: when K=3, the weight of each scale =[0.5,0.3,0.2], the diffusion parameter σ_k=[1.0,3.0,8.0], and the distance between monomers =2.5cm, the corresponding thermal coupling weight can be calculated;

[0044] The state correlation defined in S12 adopts intelligent dynamic similarity calculation:

[0045] ;

[0046] wherein, is an intelligent similarity function, is a context state vector, is a dynamic parameter vector; Adopt multi-dimensional feature similarity calculation:

[0047]

[0048] Wherein, ;

[0049] Calculation example: when = 0.92, = 0.89, = 0.05,

[0050] SOH similarity = ;

[0051] The multi-level adaptive weight fusion mechanism in S13 is as follows:

[0052] ;

[0053] Wherein, is an intelligent fusion function, which adopts a combination of weighted linear combination and nonlinear mapping; is a time-varying weight coefficient vector, which is adaptively adjusted according to the charging stage; is a feature vector of the charging stage, including stage identification of fast charging, equalization, and health assessment; Adopt a gating fusion mechanism:

[0054] ;

[0055] Wherein, is a gating function, which adjusts the weight of each relationship according to the charging stage ;

[0056] ;

[0057] Wherein, is a sigmoid activation function, is the weight matrix of the kth gate, is a bias vector;

[0058] Calculation example: when the fast charging stage = [1, 0, 0], = [0.6, 0.3, 0.1], the weight of each edge = 0.7, = 0.3, = 0.2, the fusion weight can be calculated by the gating function.

[0059] As a preferred, the hierarchical-grouping attention calculation formula adopted in S21 is as follows:

[0060] ;

[0061] wherein, is the node feature vector of the i-th battery monomer, is the node feature vector of the j-th battery monomer, is a multi-layer attention function, is an enhanced edge weight that integrates multi-dimensional information; is a set of attention parameters;

[0062] The formula of the intelligent adaptive feature aggregation in S22 is as follows:

[0063] ;

[0064] wherein, is an intelligent aggregation function that realizes feature fusion by using a gating mechanism and a residual connection;

[0065] ;

[0066] wherein, , , is a sigmoid activation function; the function represents the attention-weighted feature set of the i-th monomer neighbor node, is an adaptive aggregation parameter, including a gating weight , a bias parameter , and a residual connection coefficient, represents an element-wise product, represents vector connection, and the final node feature vector is obtained after multi-layer aggregation.

[0067] Calculation example: given conditions: the current feature of node i: , node i has two neighbor nodes, neighbor node 1: , attention weight , neighbor node 2: , attention weight , ;

[0068] Result After intelligent adaptive feature aggregation, the feature of node i is updated from [0.8, 0.6, 0.9] to [0.786, 0.512, 0.943], realizing effective fusion of neighbor information and adaptive weight adjustment.

[0069] The formula of the multi-target intelligent joint prediction output in S23 is as follows:

[0070] ;

[0071] in, As an intelligent prediction function, a multi-layer perceptron structure is used to realize SOC / SOH joint prediction; is the uncertainty estimation vector, which is used to evaluate the confidence of the prediction results; It is a set of joint prediction parameters, including prediction network weights, bias parameters and activation function parameters. Adopting a multi-task learning architecture: Contains three branches: SOC branch, SOH branch, and uncertainty branch:

[0072] Shared Feature Extraction: ;

[0073] Each branch prediction output:

[0074] ;

[0075] ;

[0076] uncertainty ;

[0077] in, 、 、 is the weight matrix of each branch, 、 、 is the bias vector of each branch, all of which belong to the joint prediction parameter set ; It is the shared layer weight matrix, which is used to map node features to the shared feature space. The weight matrix is ​​shared among all nodes. is the shared layer bias vector; ReLU is the rectified linear unit activation function; is the sigmoid activation function.

[0078] Calculation example: When input When setting the parameters It is an 8×4 matrix. Through network forward propagation calculation, the SOC prediction value is 0.85, the SOH prediction value is 0.78, and the uncertainty estimation value is 0.12.

[0079] Preferably, the formula for dynamically determining the stratification threshold in S31 based on the GNN prediction results and battery characteristics is as follows:

[0080] ;

[0081] in, It is an intelligent stratification function that determines the SOC level to which a monomer belongs based on a multi-factor assessment; is the temperature characteristic vector of the i-th monomer; Historical data, including past charging records and state change trends; Adaptive threshold vector, containing dynamic boundary parameters of each level; Multi-factor evaluation:

[0082] ;

[0083] Wherein is the SOC stratification result of the ith monomer, indicating the charging level to which the monomer belongs; P(Layer_k|feature vector) calculates the probability distribution of each charging level through the softmax function, which maps the input features to normalized probability values, ensuring that the sum of the probabilities of each level is 1.

[0084] ;

[0085] Wherein, is the weight vector of the kth layer, is the weight vector of the jth layer, is the bias of the kth layer, is the bias of the jth layer, and K is the total number of layers.

[0086] Calculation example: when = 0.65, = 0.91, = 28℃, set the historical data , the adaptive threshold vector , assuming there are 3 charging levels (K=3), and the weight parameters of each layer are:

[0087]

[0088] Through softmax probability calculation:

[0089]

[0090] After argmax operation, it is determined that the monomer belongs to the 2nd charging level.

[0091] The formula of multi-factor intelligent current optimization in S32 is as follows:

[0092] ;

[0093] Wherein, is the intelligent optimization function, which adopts a multi-objective collaborative optimization algorithm; is the fth correction factor vector of monomer i, including health correction, temperature correction and attention correction; is the constraint condition set, Check current range: 0.1A ≤ I ≤ 5A, temperature range: 0°C ≤ T ≤ 60°, voltage range: 2.5V ≤ V ≤ 4.6V; out of range limit to the boundary value; A set of multi-factor parameters for balancing charging speed, safety, and battery life; Optimization using weighted average:

[0094] ;

[0095] Where, is the optimized charging current value of the ith monomer; is the reference current determined by the hierarchical strategy The correction factor is:

[0096] ;

[0097] Where, is the fth characteristic parameter of monomer i (such as SOC, SOH, temperature, etc.), is the proportional coefficient of the fth correction factor, is the offset of the fth correction factor, with a value range is the safety constraint boundary.

[0098] Calculation example: when the monomer belongs to the 2nd charging level, the reference current = 2A, the correction factor includes: =0.9, temperature correction =1.1, attention correction = 0.95. Then:

[0099] = 2 × 0.9 × 1.1 × 0.95 = 1.881A,

[0100] After hierarchical constraint verification, the output optimized charging current value is 1.88A.

[0101] The formula for multi-level intelligent safety management and control in S33 is as follows:

[0102] ;

[0103] Where, is the intelligent safety management and control function, which realizes multi-level safety constraints and dynamic limits; is the safety state vector, which includes safety indicators such as overvoltage, overcurrent, and overtemperature; State judgment (0-normal, 1-attention, 2-warning, 3-danger):

[0104] : Graded according to voltage threshold [3.0V, 4.2V, 4.4V, 4.6V];

[0105] : Graded according to temperature threshold [10°C, 45°C, 50°C, 60°C];

[0106] : Graded according to current deviation threshold [0.1A, 0.3A, 0.5A];

[0107] For real-time feedback information, including current battery status and environmental conditions,

[0108] ;

[0109] For control parameter set,

[0110] . Where feedback risk level:

[0111] 0: normal rate of change and normal communication;

[0112] 1: slightly abnormal rate of change;

[0113] 2: significantly abnormal rate of change;

[0114] 3: severely abnormal rate of change or abnormal communication.

[0115] According to the safety level, apply the restriction coefficient:

[0116] Safety level 0: ;

[0117] Safety level 1: ;

[0118] Safety level 2: ;

[0119] Safety level 3: ;

[0120] Safety level 4: ;

[0121]

[0122] Where, is the restriction coefficient array.

[0123] As a preferred, the present application further comprises:

[0124] S4, an adaptive learning mechanism, specifically comprising the following steps:

[0125] S41, error estimation, formula as follows:

[0126] ;

[0127] S42, trigger condition judgment;

[0128] When any of the following conditions is met: the prediction error exceeds the threshold value, the error increases for a plurality of consecutive periods, and the battery pack temperature distribution is abnormal, the model update is triggered;

[0129] S43, online parameter update, the model parameters are updated in an incremental learning manner; formula as follows:

[0130] ;

[0131] wherein, is the current model parameter set, is the updated model parameter set, is the loss function based on the prediction error, is the gradient of the loss function with respect to the parameter θ; learning rate The error size is adaptively adjusted.

[0132] As a preferred embodiment, the multi-level adaptive weight fusion mechanism in S13 is adaptively adjusted according to the charging stage, the battery temperature gradient, and the state of health distribution, specifically: the physical connection relationship is emphasized in the fast charging stage; the thermal coupling relationship is adaptively emphasized in the balancing stage; and the state correlation is emphasized in the health assessment stage. The adjustment range of the weight coefficient is 0.01-0.99.

[0133] As a preferred embodiment, the enhanced multi-dimensional attention aggregation mechanism in S21 adopts a three-dimensional attention architecture of multi-head-multi-group-multi-scale, including local attention, global attention, and cross-layer attention. The number of attention heads can be configured to be 2-32, the number of attention groups can be configured to be 1-16, and the number of scales can be configured to be 1-8.

[0134] As a preferred embodiment, the intelligent adaptive threshold determination method in S31 dynamically adjusts the hierarchical threshold according to the overall state of the battery pack, historical data, and environmental conditions. The threshold adjustment range is: the high SOC layer threshold is 70%-98%, and the low SOC layer threshold is 5%-70%. The number of hierarchical layers can be configured to be 2-10.

[0135] As a preferred embodiment, the multi-factor intelligent current optimization in S32 adopts a multi-objective collaborative optimization strategy, which considers multiple objectives including charging speed, battery life, temperature control, safety constraints, and energy efficiency ratio. The optimal current distribution scheme is determined by one or more intelligent optimization algorithms such as Pareto optimal solution selection, genetic algorithm, and particle swarm optimization. The number of correction factors can be configured to be 3-20, and the correction range is 0.1-5.0.

[0136] The method also comprises a multi-scene adaptive configuration mechanism, which can automatically adjust algorithm parameters, model structures and control strategies according to different application scenarios including but not limited to household charging, commercial fast charging, mobile charging and energy storage applications, so as to realize one algorithm framework adapting to various application requirements.

[0137] The adaptive electric vehicle battery state prediction and intelligent charging system based on a graph neural network comprises:

[0138] A multi-source data acquisition layer is used to acquire multi-dimensional parameters of each single battery of a battery pack in real time, including voltage, current, temperature and internal resistance.

[0139] An intelligent graph modeling layer is used to construct an adaptive dynamic graph structure reflecting the multi-dimensional correlation relationship between single batteries.

[0140] An enhanced GNN inference layer is used to realize high-precision joint prediction of SOC / SOH / life based on a multi-level graph attention network.

[0141] An intelligent strategy generation layer is used to generate individualized, adaptive and multi-objective optimized charging strategies according to the prediction results.

[0142] A distributed intelligent execution layer is used to execute accurate charging control under multi-level safety constraints and real-time feedback control.

[0143] An online learning optimization layer is used to dynamically optimize model parameters and control strategies according to multi-dimensional feedback information.

[0144] A multi-scene configuration management layer is used to automatically adjust system configuration and operation parameters according to different application scenarios.

[0145] An intelligent safety monitoring layer is used to monitor the running state of the system in real time and provide multi-level intelligent safety protection.

[0146] A lithium battery pack is used to monitor the battery parameters collected by the data acquisition layer.

[0147] Compared with the prior art, the present application has the following beneficial effects:

[0148] (1) The adaptive electric vehicle battery state prediction and intelligent charging method and system based on a graph neural network of the present application can significantly improve the SOC / SOH prediction accuracy through intelligent multi-dimensional correlation modeling, and reduce the prediction error to below 1.5%; the differentiated intelligent charging strategy can shorten the charging time by 20%-40% under the premise of ensuring safety; and individualized charging can help to reduce the difference between single batteries by 60%-80% and improve the overall performance.

[0149] (2) The adaptive electric vehicle battery state prediction and intelligent charging method and system based on the graph neural network of the application improves the adaptability of the system to various working conditions through an online learning optimization mechanism, has good scalability, is suitable for different types and scales of battery groups, and the universality is improved by more than 90%.

[0150] (3) The adaptive electric vehicle battery state prediction and intelligent charging method and system based on the graph neural network of the application significantly expands the protection range, improves the universality and adaptability of the system through the adaptive weight fusion mechanism and the configurable parameter design, and can be applied to different types of battery groups and charging scenes.

[0151] (4) The adaptive electric vehicle battery state prediction and intelligent charging method and system based on the graph neural network of the application breaks through the limitations of traditional fixed parameters, realizes truly intelligent and personalized charging control through the improved multi-dimensional attention mechanism and adaptive threshold determination method.

[0152] (5) The adaptive electric vehicle battery state prediction and intelligent charging method and system based on the graph neural network of the application significantly improves the practicability and safety of the system through the multi-scene adaptive configuration mechanism and intelligent safety monitoring, reduces the safety risk by more than 80%, and prolongs the battery life by 15%-30%. BRIEF DESCRIPTION OF DRAWINGS

[0153] Figure 1 is the system overall architecture diagram of the application;

[0154] Figure 2 is the multi-level battery monomer relationship diagram construction flowchart of the application;

[0155] Figure 3 is the graph attention network structure schematic diagram of the application;

[0156] Figure 4 is the adaptive charging strategy generation flowchart of the application;

[0157] Figure 5 is the hardware control system architecture diagram of the application;

[0158] Figure 6 is the adaptive learning mechanism flowchart of the application. DETAILED DESCRIPTION

[0159] The technical solutions of the application will be described below in a clear and complete manner. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0160] Referring to the drawings Figure 1 The application proposes an adaptive electric vehicle battery state prediction and intelligent charging system based on a graph neural network, mainly including the following modules:

[0161] A multi-source data acquisition layer is used to acquire multi-dimensional parameters such as voltage, current, temperature, and internal resistance of each single cell of the battery pack in real time.

[0162] An intelligent graph modeling layer is used to construct an adaptive dynamic graph structure reflecting the multi-dimensional correlation between single cells.

[0163] An enhanced GNN inference layer is used to realize high-precision joint prediction of SOC / SOH / life based on a multi-level graph attention network.

[0164] An intelligent strategy generation layer is used to generate individualized, adaptive, and multi-objective optimized charging strategies according to the prediction results.

[0165] A distributed intelligent execution layer is used to execute precise charging control under multi-level safety constraints and real-time feedback control.

[0166] An online learning optimization layer is used to dynamically optimize model parameters and control strategies according to multi-dimensional feedback information.

[0167] A multi-scenario configuration management layer is used to automatically adjust system configurations and operating parameters according to different application scenarios.

[0168] An intelligent safety monitoring layer is used to monitor the system running state in real time and provide multi-level intelligent safety protection.

[0169] A lithium battery pack is used to monitor the battery parameters collected by the data acquisition layer.

[0170] Referring to the drawings Figure 2 This embodiment takes a typical 96-section lithium iron phosphate battery pack as an example to illustrate the construction process of the battery cell relationship graph in detail, and the specific steps are as follows:

[0171] Step 1: Extend the node feature vector construction;

[0172] For the i-th single cell in the battery pack, a multi-element feature vector containing basic measurements, derived calculations, context environment, and historical state is constructed:

[0173] ;

[0174] Among them, the basic measurement features , , ; the derived features include the change rates of voltage and temperature , and the internal resistance estimate value The historical features include the SOC and SOH state values at the previous time.

[0175] Step 2: Intelligent multi-dimensional edge relationship modeling

[0176] The physical topology relationship adopts an intelligent configurable weight model:

[0177] ;

[0178] In a specific implementation, for series-connected monomers i and j, For monomers within a parallel branch, ; 0 for other cases. Adaptive weight parameter Adjust dynamically according to the battery pack topology and load distribution.

[0179] The heat conduction coupling relationship adopts a multi-scale adaptive diffusion model:

[0180] ;

[0181] Consider the physical distance , multi-scale diffusion parameters and temperature field distribution to accurately model the thermal coupling relationship.

[0182] The electrochemical state correlation relationship adopts intelligent dynamic similarity calculation:

[0183] .

[0184] Step 3: Multi-level adaptive weight fusion

[0185] Adjust the weight coefficient dynamically according to the charging stage:

[0186] ;

[0187] In the fast charging stage, the physical connection weight ; In the equalization stage, the thermal coupling weight ; In the health assessment stage, the state-related weight .

[0188] Referring to the accompanying Figure 3 , this embodiment details the SOC / SOH intelligent joint prediction process of the enhanced graph attention network, and the specific steps are as follows:

[0189] Step 1: Multi-level enhanced attention aggregation mechanism

[0190] Adopt a three-dimensional attention architecture of multi-head-multi-group-multi-scale:

[0191] Local attention focuses on directly adjacent monomers:

[0192] ;

[0193] Global attention focuses on all monomers:

[0194] ;

[0195] Cross-layer attention realizes inter-layer information transmission:

[0196] ;

[0197] In a specific implementation, four attention layers are adopted, each layer has eight attention heads, and each head processes two feature groups.

[0198] Step 2: Intelligent adaptive feature aggregation;

[0199] The combination strategy of the gating mechanism and residual connection is adopted:

[0200] ;

[0201] The gating function controls the preservation and update of information, ensuring that important information is not lost;

[0202] Step 3: Multi-target intelligent joint prediction output;

[0203] Simultaneous prediction of SOC, SOH and uncertainty:

[0204] ;

[0205] The joint loss function is designed as:

[0206] ;

[0207] Experimental results show that the SOC prediction error is reduced to below 1.2%, and the SOH prediction error is reduced to below 1.8%.

[0208] Referring to the accompanying Figure 4 , this embodiment details the generation and safety control process of the multi-level adaptive charging strategy.

[0209] Step 1: Intelligent dynamic hierarchical strategy;

[0210] Dynamic threshold optimization based on reinforcement learning is adopted:

[0211] ;

[0212] Hierarchical strategy execution:

[0213] High SOC layer ( , range 70%-98%): low current charging (0.1C-0.3C);

[0214] Middle SOC layer ( ): Standard current charging (0.3C-0.8C);

[0215] Low SOC layer ( , range 5%-70%): High current charging (0.5C-1.2C);

[0216] Among them, and Real-time dynamic adjustment according to the overall state of the battery pack, historical data and environmental conditions.

[0217] Step 2: Multi-factor intelligent current optimization;

[0218] Establish a multi-objective collaborative optimization model:

[0219] ;

[0220] Correction factors include: health correction , temperature correction , attention correction , etc. The number of correction factors can be configured to 3-20, and the correction range is 0.1-5.0.

[0221] Step 3: Multi-level intelligent safety management and control;

[0222] Establish a hierarchical early warning mechanism:

[0223] ;

[0224] Safety levels include: normal level, attention level, warning level, danger level, emergency level, each level corresponds to different current limit strategy.

[0225] Experimental verification shows that this method can shorten the charging time by 25%-45%, reduce the difference between single cells by 65%-85%, and improve the charging safety by more than 80%.

[0226] Online adaptive learning mechanism based on prediction error, refer to Figure 6 , dynamically adjust the graph structure weight and model parameters according to the actual prediction error, using incremental learning method:

[0227] ;

[0228] Multi-scene adaptive configuration can automatically adjust algorithm parameters, model structure and control strategy according to different scenes such as household charging, commercial fast charging, mobile charging and energy storage applications, realizing a set of algorithm framework to adapt to multiple application requirements.

[0229] Refer to Figure 5The shown hardware control system adopts high-performance MCU as the main controller, realizes data interaction through CAN bus, uses thyristor or MOSFET switch array to realize single-stage current regulation, and is equipped with perfect safety protection circuit.

[0230] The foregoing description of specific exemplary embodiments of the application is intended to be illustrative only and is not intended to limit the application to the precise forms described. Many modifications and variations are possible in light of the above teachings without departing from the spirit or essential characteristics of the application. The exemplary embodiments are chosen and described in order to explain the principles of the application and its practical application to thereby enable others skilled in the art to best utilize the application and various embodiments with various modifications as are suited to the particular use contemplated. It is intended that the scope of the application be defined by the claims and their equivalents.

Claims

1. An adaptive electric vehicle battery state prediction and intelligent charging method based on a graph neural network, characterized in that, Comprise the following steps: S1, battery cell relationship diagram construction, specifically comprising the following steps: S11, for the battery monomer in the group, define the node feature vector; S12, define three types of edge relationship, including: physical connection relationship, thermal coupling relationship and state correlation; S13, the node feature vector and three types of edge relationship are fused by multi-level adaptive weight; S2, SOC / SOH joint prediction based on graph attention network, specifically comprising the following steps: S21, enhanced multi-dimensional attention aggregation mechanism, using hierarchical-group attention calculation; S22, intelligent adaptive feature aggregation; S23, multi-objective intelligent joint prediction output; S3, adaptive charging strategy generation, specifically comprising the following steps: S31, intelligent adaptive SOC hierarchical strategy, dynamically determine the hierarchical threshold according to the GNN prediction result and battery characteristics; S32, multi-factor intelligent current optimization; S33, multi-level intelligent safety management and control, ensure that the charging current is within the dynamic safety range; If the safety check passes, execute charging; if not, limit the current; the final current allocation feedback to the GNN prediction input; The hierarchical-group attention calculation formula used in S21 is as follows: ； wherein, is a node feature vector of the i-th battery cell, is a node feature vector of the j-th battery cell, is a multi-layer attention function, is an enhanced edge weight that integrates multi-dimensional information; is a set of attention parameters; The formula of intelligent adaptive feature aggregation in S22 is as follows: ； wherein, is an intelligent aggregation function, which realizes feature fusion by using a gating mechanism and a residual connection; represents the attention-weighted feature set of the i-th monomer neighbor node; is an adaptive aggregation parameter, including a gating weight, a forget gate parameter, and a residual connection coefficient; after multi-layer aggregation, the final node feature vector is obtained. The formula of multi-objective intelligent joint prediction output in S23 is as follows: ； wherein, is an intelligent prediction function, which is implemented by a multi-layer perception structure to jointly predict SOC / SOH; is an uncertainty estimation vector, which is used to evaluate the confidence of the prediction result; is a joint prediction parameter set, which includes prediction network weights, bias parameters and activation function parameters; The formula of dynamically determining the hierarchical threshold according to the GNN prediction result and battery characteristics in S31 is as follows: ； wherein, is the SOC tiering result for the i-th monomer, indicating the charging tier the monomer belongs to; is the intelligent tiering function, determining the SOC tier the monomer belongs to according to multi-factor evaluation; is the temperature feature vector for the i-th monomer; is the historical data, including past charging records and state change trends; is the adaptive threshold vector, containing dynamic boundary parameters for each tier; The formula of multi-factor intelligent current optimization in S32 is as follows: ； wherein, is the optimized charging current value for the i-th monomer; is the intelligent optimization function, employing a multi-objective collaborative optimization algorithm; is the f-th correction factor vector, including health correction, temperature correction, and attention correction; is the constraint condition set, containing current range, temperature limit, and safety boundary; is the multi-factor parameter set, used to balance charging speed, safety, and battery life; The formula of multi-level intelligent safety management and control in S33 is as follows: ； wherein, is an intelligent safety control function for implementing multi-level safety constraints and dynamic limits; is a safety state vector containing safety indicators including overpressure, overcurrent, and overtemperature; is real-time feedback information including current battery state and environmental conditions; is a control parameter set containing response strategies and limit thresholds for each safety level.

2. The adaptive electric vehicle battery state prediction and intelligent charging method based on graph neural network according to claim 1, characterized in that, In S11, for the i-th monomer in the battery group, the node feature vector is defined as: ； wherein, are voltage, current, temperature measurements, respectively; , are voltage and temperature rate of change; is an internal resistance estimate; is the state value of the previous time instant; The intelligent configurable weight model is used for the physical connection relationship defined in S12: ； wherein, is an intelligent connection function, using a topologically aware weight calculation method; and respectively represent the connection function and the branch function, used to identify the physical connection relationship between monomers; is an adaptive weight parameter vector, dynamically adjusted according to the battery pack topology structure; The multi-scale adaptive diffusion model is used for the thermal coupling relationship defined in S12: ； wherein, is a multi-scale thermal coupling function, is a physical distance between monomers i and j, is a multi-scale adaptive diffusion parameter vector, is a temperature field distribution vector; The intelligent dynamic similarity calculation is used for the state correlation defined in S12: ； wherein, is an intelligent similarity function, is a context state vector, is a dynamic parameter vector; The multi-level adaptive weight fusion mechanism formula in S13 is as follows: ； wherein, is a smart fusion function, which adopts a combination of weighted linear combination and nonlinear mapping; is a time-varying weight coefficient vector, which is adaptively adjusted according to the charging stage; is a charging stage feature vector, including stage identification such as fast charging, equalization, and health assessment.

3. The adaptive electric vehicle battery state prediction and intelligent charging method based on graph neural network according to claim 1, characterized in that, Also includes: S4, adaptive learning mechanism, specifically comprising the following steps: S41, prediction error evaluation, the formula is as follows: ； S42, trigger condition judgment; When any of the following conditions is met: prediction error exceeds threshold, continuous multiple period error increases, battery group temperature distribution is abnormal, trigger model update; S43, online parameter update, update model parameters in incremental learning mode; the formula is as follows: ； wherein, is the current model parameter set, is the updated model parameter set, is the loss function based on the prediction error, is the gradient of the loss function with respect to the parameters θ; learning rate is adjusted adaptively according to the error size.

4. The adaptive electric vehicle battery state prediction and intelligent charging method based on graph neural network according to claim 1, characterized in that, The multi-level adaptive weight fusion mechanism in S13 is adjusted adaptively according to the charging stage, battery temperature gradient and health state distribution, specifically: emphasize physical connection relationship in fast charging stage; adaptively emphasize thermal coupling relationship in equalization stage; focus on state correlation in health evaluation stage, the adjustment range of weight coefficient is 0.01-0.

99.

5. The adaptive electric vehicle battery state prediction and intelligent charging method based on graph neural network according to claim 1, characterized in that, The enhanced multi-dimensional attention aggregation mechanism in S21 uses a three-dimensional attention architecture of multiple heads, multiple groups and multiple scales, including local attention, global attention and cross-layer attention; the number of attention heads can be configured to 2-32, the number of attention groups can be configured to 1-16 groups, and the number of scales can be configured to 1-8.

6. The adaptive electric vehicle battery state prediction and intelligent charging method based on graph neural network according to claim 1, characterized in that, The intelligent adaptive threshold determination method in S31 dynamically adjusts the hierarchical threshold according to the overall state of the battery pack, historical data and environmental conditions, and the threshold adjustment range is: high SOC layer threshold 70%-98%, low SOC layer threshold 5%-70%; the number of hierarchical layers can be configured to 2-10 layers.

7. The adaptive electric vehicle battery state prediction and intelligent charging method based on graph neural network according to claim 1, characterized in that, The multi-factor intelligent current optimization in S32 adopts a multi-objective collaborative optimization strategy, comprehensively considers multiple objectives including charging speed, battery life, temperature control, safety constraints and energy efficiency ratio, and determines the optimal current distribution scheme through one or more intelligent optimization algorithms such as Pareto optimal solution selection, genetic algorithm and particle swarm optimization; The number of correction factors can be configured to 3-20, and the correction range is 0.1-5.

0.

8. The adaptive electric vehicle battery state prediction and intelligent charging system based on graph neural network, realizing the adaptive electric vehicle battery state prediction and intelligent charging method based on graph neural network as claimed in any one of claims 1-7, characterized in that, It includes: Multi-source data acquisition layer, for real-time acquisition of multi-dimensional parameters of each single battery of the battery pack, including voltage, current, temperature and internal resistance; Intelligent graph modeling layer, for constructing an adaptive dynamic graph structure reflecting the multi-dimensional correlation between single batteries; Enhanced GNN inference layer, for realizing high-precision joint prediction of SOC / SOH / life based on multi-level graph attention network; Intelligent strategy generation layer, for generating individualized, adaptive and multi-objective optimized charging strategies according to the prediction results; Distributed intelligent execution layer, for executing precise charging control under multi-level safety constraints and real-time feedback control; Online learning optimization layer, for dynamically optimizing model parameters and control strategies according to multi-dimensional feedback information; Multi-scene configuration management layer, for automatically adjusting system configuration and operating parameters according to different application scenarios; Intelligent safety monitoring layer, for real-time monitoring of system operation state and providing multi-level intelligent safety protection; Lithium battery pack, for monitoring battery parameters collected by the data acquisition layer.

Images