Study On The Remaining Life Prediction Method Of Lithium-ion Batteries Based On Deep Learning

In recent years,the lithium-ion battery industry has developed rapidly,and has been widely used in many fields such as new energy vehicles,energy storage systems,and consumer electronics.However,with the development,its security issues have become increasingly prominent.Accurate prediction of Remaining Useful Life(RUL)of Li-ion batteries is one of the key technologies to ensure the safe application of Li-ion batteries.At present,according to the differences in research methods,the prediction methods of RUL of Li-ion batteries are mainly divided into three categories.One is the prediction method based on the mechanism model,the second is the data-based prediction method,and the third is the fusion method containing the above two types of methods.Mechanistic model-based prediction methods have higher prediction accuracy,but are too dependent on the accuracy of the model established,and very accurate models are often difficult to obtain.Data-driven methods only need lithium battery degradation data to achieve prediction,which is widely applicable and not affected by external environment,but the RUL prediction accuracy of single data-driven forecasting methods is often limited.Fusion prediction methods can combine the advantages of both types of methods to ensure the prediction accuracy while having strong generality and robustness.Therefore,this topic has carried out data-based lithium-ion battery RUL fusion prediction methods based on data-based,combined with a variety of data-driven methods to make up for the lack of a single RUL prediction method and improve RUL prediction performance.The specific research of this article is as follows:(1)Study the decline process of lithium-ion batteries and collect decline data.This paper takes lithium iron phosphate battery as the research object,firstly introduces the theoretical knowledge of lithium battery working mechanism and basic characteristics,and analyzes the internal factors affecting the battery decline and the external reasons causing internal changes according to the working mechanism of lithium battery.Secondly,the battery experimental platform is built to test the basic performance of the battery,and after the test is completed,the battery is subjected to cyclic aging experiments to record the relevant data and study the decline process of lithium batteries.Finally,the lithium battery data set is pre-processed so that the data set can be directly used as the input of the neural network model,while avoiding poor prediction performance due to data redundancy.(2)To address the problem of poor prediction accuracy of traditional lithium battery RUL prediction methods,this paper proposes a RUL prediction method based on the Convolutional Neural Network(CNN)model.The model learns the recession characteristics of different dimensions and scales in the Li-ion battery dataset through the convolutional layer,compresses the data and parameters using the pooling layer,and selects a suitable activation function to enhance the nonlinear expression capability,and finally outputs the results by the fully connected layer.The model gets the final prediction results after continuous iteration and parameter tuning.The results show that the prediction accuracy of the convolutional neural network(CNN)-based model is significantly improved compared with the traditional RUL prediction method based on Support Vector Machines(SVM),which verifies the superiority of the data-driven prediction method.(3)To address the problems of poor long-term prediction accuracy and insufficient expression of uncertainty capability of a single RUL prediction method,this paper proposes a RUL prediction method based on Long-Short Term Memory(LSTM)model,which reduces the cumulative error and has significant long-term information prediction performance.Then we propose a RUL fusion prediction method based on Beyesian Model Averaging(BMA)that fuses multiple LSTMs.The BMA method fuses multiple LSTM sub-models according to the posterior probability,which can effectively compensate for the lack of model uncertainty expression,and at the same time the prediction performance of the model is significantly improved.The experimental results show that the prediction performance of the fused RUL prediction algorithm is significantly improved compared with that of the single RUL prediction method.