Research On Prediction Of Remaining Useful Life Of Lithium Batteries Based On Data-driven And Fusion Technology

As an important part of the new energy industry,Li-ion battery has been widely used and developed with many advantages.For the aging of Li-ion batteries,reliable Remaining Useful Life(RUL)prediction of Li-ion batteries can guarantee their safe and efficient operation and make timely warnings,etc.Therefore,the effective prediction of Li-ion battery RUL has become one of the core research contents of Li-ion battery management system.The Datadriven approach only requires battery historical data without battery internal mechanism knowledge,so this thesis uses Data-driven deep learning fusion algorithm for Li-ion battery RUL prediction,and the main contents are:Two RUL prediction models for lithium batteries based on Convolutional Neural Network(CNN)and Long Short Term Memory(LSTM)networks were developed.Historical capacity data were used as input using sliding windows,and the capacity degradation characteristics of Li-ion battery were extracted by the CNN and LSTM networks respectively,and the future capacity change was predicted to determine the RUL termination period,and the RUL was obtained according to the termination period.The experimental results of each of the two models validated the effectiveness of the deep learning model in Li-ion battery RUL prediction,and showed that the LSTM network,which is good at handling temporal data,has better performance in Li-ion battery RUL prediction.To further improve the RUL prediction accuracy,a fusion type model CNN-LSTMAttention based on hybrid network and Attention mechanism(Attention)was proposed.Using a sliding window to input the continuous historical capacity value of lithium battery with set step,CNN was used to extract spatial features from the historical capacity data,and then the LSTM network was used to extract temporal features from the CNN output vector,and finally the attention mechanism was used to capture the key spatio-temporal characteristics of the LSTM output vector to give the single-step forward capacity prediction results,and successive iterative operations were performed to obtain the future temporal capacity,and then the RUL prediction values were obtained,and Dropout was added to the network to avoid the overfitting phenomenon.The experimental results showed that the CNN-LSTM-Attention network can make accurate predictions of Li-ion battery RUL,and the comparison with two single models and a hybrid model without the Attention mechanism showed that the CNN-LSTM-Attention network has higher accuracy and generalization ability.The uncertainty analysis of RUL results can provide decision advice for the maintenance of lithium batteries,etc.In order to obtain the uncertainty of the RUL prediction results for lithium batteries,the Monte Carlo(MC)Dropout method combined with CNN-LSTMAttention network was proposed to obtain the uncertainty of RUL prediction results.Based on the fact that the Bayesian network after variational inference can give the probability distribution of the results,it is proposed that adding the Dropout technique can make the CNNLSTM-Attention network approximate the Bayesian network,and then combined with MC sampling to obtain the uncertainty of the RUL prediction results.After comparing and analyzing the experimental results,the effectiveness of MC Dropout for quantifying the uncertainty of prediction results was verified,and the results from the uncertainty quantification verified that the CNN-LSTM-Attention proposed in this project has stronger robustness and certainty than other networks.