CN-121995231-A - Battery life prediction model, system and readable storage medium

CN121995231ACN 121995231 ACN121995231 ACN 121995231ACN-121995231-A

Abstract

The application discloses a battery life prediction model, a system and a readable storage medium, six modules of data processing, interactive coding, multi-scale fusion, global prediction, trend modeling and layered fusion are combined, and each module is used for realizing high-precision battery life prediction in a cooperative manner. The interactive coding module adopts a double-shaft attention mechanism to mine the association and timing rules of the features and output the enhanced features. The multi-scale fusion module extracts and fuses the long-term trend and the local fluctuation feature through a double-branch structure, and the global prediction module generates a global prediction result through feature aggregation, self-adaptive weighted fusion and regression mapping. The trend modeling module separates the core performance index sequence, captures the battery aging rule through the dual-path network, and the layering fusion module fuses two types of prediction results by means of the gating network to avoid single mode deviation. The modules cooperatively improve the data utilization rate, and the prediction precision and reliability are obviously improved.

Inventors

Tian Dingwei
YANG HAO
YU HONGJIANG
LIU HONGYONG

Assignees

江苏正力新能电池技术股份有限公司

Dates

Publication Date: 20260508
Application Date: 20260212

Claims (10)

1. A battery life prediction model, comprising: The data processing module is used for carrying out feature extraction and pretreatment on the battery history circulation data to construct an input sequence, wherein the features at least comprise relaxation voltage features, temperature features and circulation-level integral features; The interactive coding module adopts a double-shaft attention mechanism to sequentially perform interactive attention calculation of feature dimension and time sequence dependent coding of time dimension on the input sequence, and outputs enhanced feature representation; The multi-scale fusion module comprises a time sequence attention branch and a dynamic multi-scale convolution branch which are parallel, and is used for extracting and fusing long-term trend characteristics and local fluctuation characteristics of the enhanced characteristic representation to generate a comprehensive characteristic representation; The global prediction module is used for carrying out feature aggregation on the comprehensive feature representations, carrying out weighted fusion on the feature representations from different branches based on self-adaptive weights, and further generating a global prediction result through regression mapping; the trend modeling module is used for separating a core performance index sequence from the input sequence, carrying out attenuation trend prediction through a dual-path network structure based on the core performance index sequence, and generating a trend prediction result, wherein the core performance index sequence at least comprises four core performance indexes of charge capacity, discharge capacity, charge energy and discharge energy; and the hierarchical fusion module is used for carrying out dynamic weighted fusion on the global prediction result and the trend prediction result through a trend fusion gating network to generate a final battery life prediction value.
2. The model of claim 1, wherein the interactive coding module comprises: The feature interaction attention unit is used for carrying out dimension rearrangement on the input sequence, applying a multi-head self-attention mechanism on feature dimensions, calculating association weights among different features and generating an intermediate feature matrix containing feature interaction information; And the time sequence dependent coding unit is used for applying a transducer coder to the intermediate feature matrix in the time dimension, capturing the long-term time sequence dependent relation between the historical cyclic data and outputting the enhanced feature representation.
3. The model of claim 1, wherein the time-series attention branches comprise: the time sequence dependent encoder is a time dimension transform encoder and is used for encoding the enhanced feature representation obtained through the feature interaction attention module processing so as to model the long-term time sequence dependent relation of the battery cycle sequence and output the time sequence dependent enhanced representation.
4. A model according to claim 3, characterized in that the dynamic multi-scale convolution branches comprise: The lightweight controller network is used for dynamically generating weight coefficients respectively corresponding to a plurality of preset convolution kernel sizes according to the input characteristic representation; The multi-scale parallel convolution layer is composed of a plurality of one-dimensional convolution layers with different convolution kernel sizes, is used for extracting local time sequence modes of the enhancement features in parallel and outputting a plurality of convolution features; And the weighted fusion layer is used for carrying out weighted summation on the plurality of convolution features according to the weight coefficient generated by the lightweight controller network to generate dynamic convolution features.
5. The model of claim 4, wherein the multi-scale fusion module further comprises a branch fusion unit; The branch fusion unit adopts a self-adaptive attention mechanism to dynamically calculate a first fusion weight of the time sequence dependency enhancement representation and a second fusion weight of the dynamic convolution feature, and performs weighted fusion on the time sequence dependency enhancement representation and the dynamic convolution feature according to the first fusion weight and the second fusion weight to generate the comprehensive feature representation.
6. The model of claim 1, wherein the dual path network structure comprises: The instantaneous state modeling path is used for extracting an index value of the latest time step in the core performance index sequence and generating a first trend feature through a first multi-layer perceptron; The time sequence evolution modeling path is used for carrying out one-dimensional rolling and pooling operation on the complete core performance index sequence, extracting time sequence evolution characteristics and generating second trend characteristics through a second multi-layer perceptron; and the trend internal fusion unit is used for dynamically fusing the first trend characteristic and the second trend characteristic through a trend fusion gating network to generate the trend prediction result.
7. The model of claim 6, wherein the calculation formula corresponding to the trend-fusion-gating network fusion-generated trend prediction result is: Wherein, the The applied fusion weights are calculated for the trend prediction results, The processing result of the time characteristics by the gating network is fused for trend, As a first trend feature of the device, As a feature of the second trend, Is a trend prediction result.
8. The model of claim 1, wherein the data processing module constructs the input sequence through a sliding window mechanism; the sliding window mechanism is as follows: Sequentially intercepting continuous characteristic data of a plurality of loops from a loop sequence of a full life cycle of a battery to form a training sample, and taking a battery performance index of the loops after a specified future interval as a prediction target of the training sample; And generating a training sample sequence with time sequence relevance by sliding a window with fixed length along the cyclic sequence, and taking the training sample sequence as the input sequence.
9. A battery life prediction system, characterized in that the system is integrated with a battery life prediction model according to any one of claims 1 to 8.
10. A readable storage medium having stored thereon parameters of a battery life prediction model, characterized in that the battery life prediction model is a model as defined in any one of claims 1 to 8.

Description

Battery life prediction model, system and readable storage medium Technical Field The present application relates to the field of battery analysis technology, and more particularly, to a battery life prediction model, a system, and a readable storage medium. Background The lithium ion battery is used as the core of the modern energy storage technology, is widely applied to the key fields of electric automobiles, energy storage systems, consumer electronics and the like, and the accurate prediction of the residual service life (RUL) of the lithium ion battery has important engineering value and economic significance for guaranteeing the safe operation of the system, optimizing the operation and maintenance cost and improving the user experience, and particularly realizes the full life cycle performance prediction in the early stage of the use of the battery, and is very important for the intelligent decision making and preventive maintenance of a battery management system. Currently, the battery remaining life prediction technology is mainly divided into three routes based on a physical model, a data driving method and a hybrid method, and although certain progress is made, significant technical limitations exist. Although the physical modeling method based on the electrochemical mechanism has good physical interpretability, the problems of complex parameter identification, excessive calculation load and insufficient individual difference adaptability are faced, and the practical application requirements are difficult to meet. The data driving method is subjected to evolution from statistic learning to deep learning, but has systematic defects that the feature extraction lacks electrochemical mechanism support, mostly adopts simple statistical features with fuzzy physical meaning, ignores key electrochemical information such as relaxation process and the like, has low data utilization efficiency, causes a large amount of information loss by adopting a single cycle and single feature vector mode, and also lacks an effective time sequence data enhancement strategy, so that training samples are insufficient and the advantage of deep learning cannot be exerted. The existing deep learning method mostly adopts a general time sequence architecture, lacks interaction modeling capability among features, is difficult to fuse multi-scale time sequence information such as long-term trend and short-term fluctuation of battery attenuation, has insufficient field knowledge fusion, and has difficulty in meeting actual engineering requirements in model generalization capability and prediction accuracy. Therefore, a new battery life prediction method is needed to solve the defects of the prior art and realize high-precision and high-reliability battery residual service life prediction. Disclosure of Invention The application provides a battery life prediction model, a system and a readable storage medium, which realize high-precision and high-reliability battery residual service life prediction through the cooperative work of six modules of data processing, interactive coding, multi-scale fusion, global prediction, trend modeling and layered fusion. A battery life prediction model comprising: The data processing module is used for carrying out feature extraction and pretreatment on the battery history circulation data to construct an input sequence, wherein the features at least comprise relaxation voltage features, temperature features and circulation-level integral features; The interactive coding module adopts a double-shaft attention mechanism to sequentially perform interactive attention calculation of feature dimension and time sequence dependent coding of time dimension on the input sequence, and outputs enhanced feature representation; The multi-scale fusion module comprises a time sequence attention branch and a dynamic multi-scale convolution branch which are parallel, and is used for extracting and fusing long-term trend characteristics and local fluctuation characteristics of the enhanced characteristic representation to generate a comprehensive characteristic representation; The global prediction module is used for carrying out feature aggregation on the comprehensive feature representations, carrying out weighted fusion on the feature representations from different branches based on self-adaptive weights, and further generating a global prediction result through regression mapping; the trend modeling module is used for separating a core performance index sequence from the input sequence, carrying out attenuation trend prediction through a dual-path network structure based on the core performance index sequence, and generating a trend prediction result, wherein the core performance index sequence at least comprises four core performance indexes of charge capacity, discharge capacity, charge energy and discharge energy; and the hierarchical fusion module is used for carrying out dynamic weighted fusion on the glob