Data-driven Time Series Prediction Of Lithium Battery Health Statu

With the advancement of technology,time series prediction methods have gradually developed from single variable linear analysis to multivariate nonlinear analysis,and have been widely applied in various research fields.Such as battery state,future weather,environmental conditions,and traffic flow.However,due to the increasing amount of input data and the rising difficulty of task requirements,traditional prediction models are unable to meet current demands.Therefore,this thesis aims to improve the accuracy and stability of time series prediction of battery health state by adopting machine learning and deep learning methods.The main research contents are as follows:(1)Battery Health State Prediction Model Based on Heterogeneous Ensemble Algorithm.Based on a comprehensive comparison of the predictive performance of various tree model ensemble algorithms,a heterogeneous double-layer Stacking ensemble model was designed to improve the overall predictive performance of the heterogeneous ensemble model by combining multiple tree models,linear models,and nonlinear models.(2)Building Battery Health State Prediction Model Based on Deep Neural Networks.To explore the solution of deep network models for time series problems,this article constructs a model network in stages from the perspective of convolutional structure,recurrent neural network functionality,and multi-receptive field,and trains and validates the model performance step by step,demonstrating the gain ability of the model structure.(3)Building Battery Health Prediction Model Based on Linear Feature Fusion Neural Network.To address the issue of long-term prediction drift encountered in the experimental process,this article explores the feasibility of a linear feature fusion network model.Bi GRU neural network and Self-Attention mechanism are used to enhance the model’s ability to capture long-term features.By parallelizing the linear network branches,the responsiveness of the model autoregressive target is achieved,thereby enhancing the model’s responsiveness to input amplitude changes.The whale optimization algorithm is used to optimize the model parameters,and the effectiveness of each module and the superiority of the proposed model are verified through ablation study.