Research On Data-driven Fusion Prediction Methods Of Lithium-ion Battery Remaining Useful Life

Lithium-ion battery is widely used in various elect ronic products,and gradually expanded to electric vehicles,aircraft,spacecraft and other fields.However,the lithium-ion battery also pose safety risks.The research on the prediction of remaining useful life(RUL)of lithium-ion batteries is the key technology to ensure the safe application of lithium-ion batteries.The prediction accuracy of RUL based on mechanism model is high,but it is difficult to obtain general accurate mechanism model.If only using one data-driven RUL prediction method for lithium battery,the prediction performance is limited.Therefore,this thesis studies the data-driven fusion prediction method for the RUL of lithium-ion batteries based on the historical degradation data of batteries,which is independent of the degradation mechanism model of batteries under specific environmental conditions and working conditions,to improve the versatility and prediction accuracy of the RUL prediction method.In view of some problems in the current RUL method of lithium-ion batteries,the fusion prediction method of RUL for lithium-ion batteries is studied from the following three aspects:(1)For the problem that particle filter(PF)prediction method relies on battery empirical degradation model and can not obtain real observation data in multi-step process.In this thesis,we propose a method based on Bayesian regression and dynamic state estimation for RUL fusion prediction of Li-ion batteries.We use the relevance vector machine based on Bayesian regression theory to build a state estimation model independent of battery empirical degradation model,and introduce the auto regressive(AR)model to approximate the observed value of PF method by the long-term battery capacity degradation trend prediction value.So the parameters of state esti mation model are updated dynamically.The experimental results show that the prediction accuracy of the fusion method is slightly higher than that of the classical PF method without the dependence of the complex empirical degradation model of the battery.(2)For the prediction method of RUL based on RVM machine learning model,the current RVM model is mainly single kernel,and the selection of kernel function is subjective.In this thesis,a fusion prediction method of RUL for lithium battery based on optimized multi-kernel RVM model is proposed,the most representative radial basis function(also known as gaussian kernel function),linear kernel function,polynomial kernel function and sigmoid kernel function of RVM are selected respectively.The fruit fly optimization algorithm(FOA)is introduced to optimize the linear combination of different kernel functions,so as to integrate the comprehensive expression ability of multiple kernel functions and effectively improve the prediction accuracy of the model.The experimental results show that the RUL fusion prediction method based on the optimized multi-kernel RVM model improves the RUL prediction accuracy compared with the RUL prediction method based on the single kernel RVM model and the support vector regression(SVR)model.(3)For the poor long-term prediction accuracy of the RUL prediction method based on the shallow machine learning model and the lack of uncertainty management of some methods,a prediction method based on the long short-term memory network(LSTMN)deep learning model is proposed,reduced cumulative error and improved prediction accuracy.In the end,a RUL fusion prediction method based on bayesian model averaging(BMA)integrating multiple LSTMN deep learning models is proposed.BMA alg orithm with uncertainty expression ability is used to fuse multiple LSTMN models,which not only makes up for the lack of uncertainty expression in deep learning model,but also further improves the RUL prediction accuracy of lithium battery based on deep learning model.The experimental results show that the prediction accuracy of RUL fusion prediction method based on deep learning model is much higher than that based on shallow machine learning model.