CN-122020212-A - Intelligent optimization method for SOH estimation of cross-chemistry system battery

CN122020212ACN 122020212 ACN122020212 ACN 122020212ACN-122020212-A

Abstract

The invention provides an intelligent optimization method for SOH estimation of a cross-chemical system battery, belongs to the field of health state estimation of an electrochemical energy storage system, and is used for solving the problems of strong data dependence and poor cross-system generalization of the traditional method. The method comprises the steps of preprocessing time sequence data of a cross-chemical system battery to obtain a robust feature matrix, building a gated convolutional neural network to extract degradation features, combining self-adaptive optimization and super-parameter search optimization models, learning common initial parameters through a cross-system element training task to realize adaptation of a small number of samples in a target domain, improving performance through model weight reduction and working condition calibration, clustering common parameters, and finishing deployment in a layered fine tuning mode. The data dependence is reduced, the generalization capability of a cross-system, the robustness of complex working conditions and the deployment efficiency are improved, and the method is suitable for the SOH real-time estimation of a cross-chemical system battery.

Inventors

LIN JIAZHEN
XUE ZHONGSHENG
JIA PING
WANG WENXIN
WANG RUI
SONG YE

Assignees

三峡巴州若羌能源有限公司
中国长江三峡集团有限公司

Dates

Publication Date: 20260512
Application Date: 20251229

Claims (10)

1. An intelligent optimization method for cross-chemistry system battery SOH estimation is characterized by comprising the following steps: S1, collecting electrochemical parameter time sequence data of a cross-chemical system battery, intercepting an effective segment to construct a feature matrix, eliminating numerical value difference through cross-system normalization processing, and acquiring the feature matrix by combining noise filtering and outlier correction; s2, constructing a gating convolutional neural network, extracting degradation characteristics of a cross-system battery, and outputting an SOH preliminary estimated value; s3, acquiring an optimal parameter combination by adopting a self-adaptive parameter optimization algorithm and a regularization mechanism and combining a super-parameter intelligent search mechanism, and training and optimizing a deep learning basic model; s4, constructing a cross-chemical system meta-training task, optimizing and learning a cross-system commonality initial parameter through task self-adaptive parameter adjustment and meta-loss, and adapting a small number of samples in a target domain based on the initial parameter; S5, performing parameter reduction, parallel computing configuration and data quantization processing on the optimized model, and calibrating the model by adopting a distribution offset measurement method, adaptive screening threshold adjustment and multi-mode feature fusion processing; s6, clustering and caching cross-system commonality parameters according to chemical system characteristics, adjusting system specificity parameters during target domain deployment, and completing model deployment by adopting a self-adaptive convergence termination mechanism.
2. The intelligent optimization method for SOH estimation of a cross-chemistry system battery according to claim 1, wherein the step S1 specifically comprises the following steps: S11, collecting time sequence data of voltage, current, temperature, charging capacity and increment capacity of a battery crossing a chemical system, wherein the increment capacity is obtained by differential calculation of the charging capacity and the voltage by referring to a formula (1), the multidimensional time sequence data is organized into a feature matrix according to a formula (2), and the formula is as follows: (1); Wherein, the Indicating the incremental capacity of the i-th second, And Charge capacities of i-th second and i-1 th second, respectively; And The voltages for the ith second and the i-1 th seconds respectively; (2); Wherein, the Voltage, current, temperature, charge capacity and delta capacity for the ith second, respectively; The number of the cut fragments; S12, carrying out dimensionless treatment on the characteristics of the voltage and the charging capacity by referring to normalization formulas (3) and (4) to eliminate numerical value differences of different chemical systems, wherein the formulas are as follows: (3); (4); Wherein, the Is the non-dimensionalized ith second voltage; The charge capacity is the i second charge capacity after dimensionless; And Respectively corresponding to rated voltage and rated capacity of the chemical system battery; s13, noise filtering the normalized features by referring to the sliding window mean filtering of the formula (5), and calculating the median of the feature sequence by referring to the formulas (6) and (7) respectively Identifying abnormal values with the absolute deviation median MAD, replacing the identified abnormal values by adopting a sliding window mean value by referring to a formula (8), wherein the formula is as follows: (5); Wherein, the Is the filtered i second characteristic value; The original characteristic value of the kth second; (6); (7); (8); Wherein, the And the characteristic value after the abnormal value replacement.
3. The intelligent optimization method for SOH estimation of a cross-chemistry system battery according to claim 1, wherein step S2 specifically comprises the steps of: s21, the gating convolutional neural network is formed by sequentially connecting a convolutional layer, an activation function layer, a gating layer, a maximum pooling layer and a linear layer in series, wherein a model is input into a robust feature matrix, an SOH estimated value is output through a formula (9), and the formula is as follows: (9); Wherein, the The current capacity of the battery; is the rated capacity of the battery; representing a mapping function of GCNNs models, and converting an input feature matrix into a current capacity predicted value through each layer of operation of the models; S22, after zero vector filling is carried out on the robust feature matrix, local features are extracted from a plurality of parallel one-dimensional convolution kernels through a formula (10), convolution feature tensors are output, and the formula is as follows: (10); Wherein, the Output eigenvalue at the ith second for the nth convolution kernel; the jth weight parameter being the nth convolution kernel; A 5-dimensional feature vector of the (i+j-1) th row of the feature matrix after filling; A bias vector that is the nth convolution kernel; S23, sequentially applying a ReLU function of a formula (11) and a tanh function of a formula (12) to the convolution characteristic tensor to carry out nonlinear enhancement, wherein the formula is as follows: (11); Wherein, the The characteristic value is output by the convolution layer; Is the characteristic value after ReLU activation; (12); Wherein, the The characteristic value after the tanh is activated; S24, calculating candidate characteristic tensors and gating signal tensors by respectively referring to formulas (13) and (14) through two independent convolution layers, multiplying the candidate characteristic tensors and the gating signal tensors element by element through a formula (15) to perform characteristic screening, wherein the formula is as follows: (13); (14); Wherein, the And Respectively representing the operation of two independent convolution layers; candidate tensors for all potentially valid features; The gating signal tensor is used for controlling the passing rate of the candidate characteristic; (15); Wherein, the Element values of the filtered feature tensor on the ith convolution kernel, the jth time step and the kth feature dimension are obtained; S25, performing time dimension compression and key feature reinforcement on the feature tensor after screening by adopting a maximum pooling layer reference formula (16), wherein the formula is as follows: (16); Wherein, the Element values of the feature tensor after pooling in the ith convolution kernel, the jth time step and the kth feature dimension; S26, expanding the pooled characteristic tensor into a one-dimensional vector through a formula (17), mapping the one-dimensional vector into a predicted value of the current capacity of the battery through a linear layer, and calculating to obtain an SOH estimated value through a formula (9) in combination with the rated capacity, wherein the formula is as follows: (17); Wherein, the Predicting the current capacity of the battery; A weight matrix that is a linear layer; Is the bias vector of the linear layer.
4. The intelligent optimization method for SOH estimation of a cross-chemistry system battery according to claim 1, wherein step S3 specifically comprises the steps of: s31, adopting a mean square error as a basic loss function, and referring to a formula (18) by a calculation formula, wherein the formula is as follows: (18); Wherein, the Is the original loss function value; all trainable parameters for GCNNs models; For training sample number; SOH predicted value for the mth sample; SOH true value for the mth sample; S32, L2 regularization is introduced, overfitting is restrained through a formula (19), and weight attenuation in the parameter updating process is conducted through a formula (20); (19); Wherein, the Adding a regularized loss function value; Is a regularization coefficient; Total number of trainable parameters for the model; Is the i-th trainable parameter; (20); Wherein, the The parameter value after the t+1st iteration is the parameter value; the parameter value is the parameter value after the t-th iteration; is the learning rate; gradient of original loss function under the parameter of the t iteration; s33, updating model parameters by adopting an Adam optimization algorithm, calculating gradients by using a formula (21), updating first moment estimation by using a formula (22), updating second moment estimation by using a formula (23), correcting moment estimation deviation by using formulas (24) and (25), and finally updating parameters by using a formula (26), wherein the formulas are as follows: (21); Wherein, the Gradient vector for the t-th iteration; (22); Wherein, the A first moment estimated value for the t-th iteration; a first moment estimated value for the t-1 th iteration; the first-order moment attenuation rate is used for smoothing the gradient and reducing gradient fluctuation; (23); Wherein, the A second moment estimated value for the t-th iteration; a second moment estimated value for the t-1 th iteration; the second moment attenuation rate; is the element-by-element square of the gradient vector, used to estimate the variance of the gradient; (24); (25); Wherein, the The first moment estimated value is corrected; The corrected second moment estimated value; t is the iteration number; (26); Wherein, the Avoiding denominator to be 0 for an extremely small constant; Is the initial learning rate; s34, searching the optimal parameter combination by using a super-parameter searching mechanism with the aim of minimizing the root mean square error of the cross-chemical system verification set.
5. The intelligent optimization method for SOH estimation of a cross-chemistry system battery according to claim 1, wherein step S4 specifically comprises the steps of: S41, selecting a plurality of cross-chemistry system battery aging data sets as meta-training data sets, and dividing the multi-chemistry system battery aging data sets into meta-training subsets and meta-verification subsets according to proportion; s42, the meta-training process comprises an inner loop and an outer loop, and model initial parameters are optimized through double loop iteration, wherein: In the adaptation of the internal circulation task, the initial parameters of the gating convolutional neural network are set as the current external circulation iteration parameters, a loss function is calculated based on an adaptation sample subset, and task specific parameters are obtained through one-time gradient descent update of a formula (28); In the external circulation element optimization, a loss function is calculated on an evaluation sample subset based on task specific parameters, and a cross-system difference penalty term is added in the element loss function by referring to a formula (30), wherein the cross-system difference is quantized by a formula (31), the Wasserstein distance is calculated, and the initial parameters of the model are updated by adopting a formula (32) through random gradient descent; The formula is as follows: (28); Wherein, the Adapting the learning rate for internal circulation; (29); Wherein, the The method comprises the steps of supporting a mean vector of the kth dimension characteristic of the set for the ith task; The mean value vector of the kth dimension characteristic of the set is supported for all meta-training tasks; (30); Wherein, the Is a value of a meta-loss function; Training the number of tasks for the element; is a penalty coefficient; The Wasserstein distance between the source domain feature distribution and the target domain feature distribution is used for quantifying the difference of the cross-system feature distribution, and punishing the parameter updating excessively biased to a single system; (31); Wherein, the Is the source domain feature distribution With target domain feature distribution Is a set of joint probability distributions; is a source domain feature sample; A target domain feature sample; Euclidean distance for samples x and y; the smallest possible value; (32); Wherein, the The learning rate of the external circulation element; S43, selecting a cross-chemical system data set different from the meta-training data set as a target domain data set, using the optimal initial parameters obtained by meta-training, performing gradient descent fine tuning based on a small amount of adaptation samples of the target domain, and obtaining model parameters of the adaptation target domain through a formula (28).
6. The intelligent optimization method for SOH estimation of a cross-chemistry system battery according to claim 1, wherein step S5 specifically comprises the steps of: S51, based on a super-parameter optimizing result, preserving a core convolution kernel with high characteristic contribution degree, removing a redundant convolution kernel, simultaneously splitting the input characteristics of a model into multiple paths of parallel calculation branches, synchronously carrying out characteristic processing, carrying out 8-bit integer quantization on model weight parameters and activation values, calculating reference formulas (33) and (34) of quantization scaling factors, and carrying out quantization operation reference formulas (35) and (36) and inverse quantization reference formulas (37) and (38) during reasoning, wherein the formulas are as follows: (33); (34); Wherein, the Quantization scaling factors for the weight parameters; A quantization scale factor for the activation value; is the maximum value of the weight parameter; 127 is the maximum positive value of the int8 type; (35); (36); Wherein, the The quantized weight parameters are obtained; the quantized activation value; (37); (38); Wherein, the Is the weight parameter after inverse quantization; is the activation value after inverse quantization; S52, referring to a formula (39), quantifying the distribution offset degree of the input features and the meta-training features through KL divergence, dynamically adjusting a gating layer feature filtering threshold value through a formula (40) according to the KL divergence value, introducing a battery surface temperature gradient feature, fusing the battery surface temperature gradient feature with a core feature through a lightweight full-connection network after splicing, and correcting the extreme working condition precision, wherein the formula is as follows: (39); Wherein, the Probability distribution for the current input feature; Training the probability distribution of the cross-system characteristics for the elements; (40); Wherein, the The threshold value of the gate control layer is adjusted; and the maximum KL divergence value is counted for the meta-training stage.
7. The intelligent optimization method for SOH estimation of a cross-chemistry system battery according to claim 1, wherein step S6 specifically comprises the steps of: S61, clustering optimal initial parameters based on core electrochemical parameters of each chemical system, and caching a clustered cross-system commonality parameter subset in the local of the edge equipment; s62, loading a corresponding common parameter subset to initialize a model during target domain deployment, and fine-tuning system specific parameters of the model; S63, when the RMSE and the fluctuation of the verification set after the continuous multi-round fine tuning are smaller than the threshold value, the fine tuning is terminated.
8. The intelligent optimization method for SOH estimation across chemical systems according to claim 5, wherein in step S42: In the meta training process, the internal circulation adaptive learning rate is dynamically adjusted according to cosine similarity, and a larger learning rate is adopted when the similarity is high, and a smaller learning rate is adopted when the similarity is low.
9. The intelligent optimization method for SOH estimation of a cross-chemistry system battery according to claim 6, wherein the method comprises the following steps: in the step S51, the multi-path parallel computing branch is divided into three paths, namely voltage and current characteristics, temperature and charging capacity characteristics and increment capacity characteristics; In step S52, the temperature gradient of the battery surface is obtained by collecting temperature data of a plurality of designated areas of the battery surface, and the temperature gradient is fused with the core characteristics after normalization.
10. The intelligent optimization method for SOH estimation of a cross-chemistry system battery according to claim 7, wherein the method comprises the following steps: The cross-system commonality parameters comprise a gating convolutional neural network convolutional kernel weight and a gating layer threshold; The system specificity parameter comprises two layers of convolution kernel weights after the model, full-connection layer weights and offsets; the fine tuning adopts Adam algorithm, and the fine tuning learning rate is positively correlated with the meta-training internal circulation learning rate.

Description

Intelligent optimization method for SOH estimation of cross-chemistry system battery Technical Field The invention belongs to the field of health state estimation of an electrochemical energy storage system, and particularly relates to an intelligent optimization method for SOH estimation of a cross-chemical system battery. Background The accurate estimation of the battery health state is a core link of the performance and reliability of an electrochemical energy storage system, and along with the energy transformation promotion, novel chemical systems such as a sodium ion battery, a vanadium flow battery and the like continuously emerge, and the traditional SOH estimation method gradually fails. The traditional deep learning method relies on large-scale labeling data training, but the novel battery aging experiment is high in cost and difficult in data acquisition, the characteristic difference of batteries of different chemical systems is obvious, the model trained by a single system is poor in mobility and easy to generate negative migration, the traditional method needs to manually design a complex feature extraction flow and is difficult to adapt to non-fixed input such as random charging fragments, and the performance of the traditional method is greatly reduced due to limited data of a newly deployed battery system. The complex working condition in the practical application is easy to cause characteristic distribution deviation, the robustness of the existing model is insufficient, and the edge deployment efficiency is low due to the complex model structure. In the prior art, the migration learning method has poor adaptation flexibility due to freezing network parameters and is mostly based on single data set pre-training, the meta learning method supports small sample adaptation but is limited to a single chemical system, the problem of cross-system characteristic difference cannot be solved, the fusion degree of an edge AI optimization technology and a battery state estimation model is insufficient, the precision and the instantaneity are difficult to balance, and therefore, the SOH estimation technology which has strong cross-chemical system adaptation, stable precision under small samples, maintains target performance under complex working conditions and can be deployed at high efficiency is urgently needed. Disclosure of Invention The invention mainly aims to provide an intelligent optimization method for SOH estimation of a cross-chemical system battery, which solves the problems of strong heterogeneity of characteristic distribution, low estimation precision and poor instantaneity caused by algorithm design defects under a passive domain due to the difference of cross-system electrochemical mechanisms. In order to solve the technical problems, the technical scheme adopted by the invention is that an intelligent optimization method for SOH estimation of a cross-chemical system battery comprises the following steps: S1, collecting electrochemical parameter time sequence data of a cross-chemical system battery, intercepting an effective segment to construct a feature matrix, eliminating numerical value difference through cross-system normalization processing, and acquiring the feature matrix by combining noise filtering and outlier correction; s2, constructing a gating convolutional neural network, extracting degradation characteristics of a cross-system battery, and outputting an SOH preliminary estimated value; s3, acquiring an optimal parameter combination by adopting a self-adaptive parameter optimization algorithm and a regularization mechanism and combining a super-parameter intelligent search mechanism, and training and optimizing a deep learning basic model; s4, constructing a cross-chemical system meta-training task, optimizing and learning a cross-system commonality initial parameter through task self-adaptive parameter adjustment and meta-loss, and adapting a small number of samples in a target domain based on the initial parameter; S5, performing parameter reduction, parallel computing configuration and data quantization processing on the optimized model, and calibrating the model by adopting a distribution offset measurement method, adaptive screening threshold adjustment and multi-mode feature fusion processing; s6, clustering and caching cross-system commonality parameters according to chemical system characteristics, adjusting system specificity parameters during target domain deployment, and completing model deployment by adopting a self-adaptive convergence termination mechanism. In a preferred embodiment, the step S1 specifically includes the following steps: S11, multi-dimensional time sequence data acquisition and fragment interception Firstly, collecting real-time operation data of batteries of different chemical systems such as lithium/sodium/vanadium flows, wherein the real-time operation data comprise voltage, current, temperature, charging capacity and increment capacity