A Research Of Remaining Useful Life Prediction Of Zinc-Ion Batteries Based On Machine Learning

Rechargeable aqueous zinc-ion batteries are widely attractive for large-scale energy storage technologies owing to their safe and environmentally friendly electrolytes.However,this does not mean that zinc-ion batteries do not have potential safety hazards.The predicting of the remaining useful life（RUL）of the battery is a key technology to ensure their further safe use.With the rise of machine learning,more and more studies have begun to use machine learning to predict the RUL of batteries,but it is still difficult to accurately predict the RUL of each battery,because the degradation process of the battery is very complex,and the chemical reaction inside the battery is random,so that the degradation trend of each battery is different.In addition,the research on battery RUL prediction is concentrated in the field of lithium-ion batteries,and there is no relevant report in the field of zinc-ion batteries.In this thesis,aiming at the accurate prediction of RUL of zinc-ion batteries,the related researches are carried out around the problems of inaccurate prediction and poor generalization of conventional machine learning models and accumulated errors in recursive prediction.The improvement of the model structure and the research on the optimization method are mainly discussed.The main research work is as follows:（1）Aiming at the problem of poor consistency of charge and discharge data,the fabrication process and data acquisition method of zinc-ion battery were studied.Through repeated experiments and research,each step in the battery manufacturing process is optimized to improve the poor consistency of hand-made batteries.Then keep the battery testing facility as consistent as possible to reduce variability in data collection.The charge-discharge data of multiple batches of batteries show that these optimization methods can effectively improve the problem of poor consistency.（2）Aiming at the problems of inaccurate prediction and poor generalization of zincion battery RUL by conventional machine learning models,an asymmetric encoderdecoder model based on attention mechanism is proposed.Inspired by the phenomenon that the biological brain has asymmetric structure,this thesis constructs an asymmetric neural network model by making the encoder and decoder structures different.With the introduction of attention mechanism,the encoder-decoder model can focus on the input information that is helpful for prediction,and applying dropout technology to the decoder can improve the generalization performance of the model.Finally,Bayesian optimization is used to find approximately optimal hyperparameters for this computationally expensive model.The experimental results show that the improved encoder-decoder model can more accurately predict the RUL of zinc-ion batteries,and the generalization performance is better than that of the conventional machine learning model.（3）Aiming at the problem of accumulative error in the improved encoding-decoding model,a fusion prediction method of zinc-ion battery RUL based on Gaussian process regression and improved encoding-decoding model is proposed.In order to enable Gaussian process regression to accurately predict the battery capacity near the training set,a new kernel function is designed in this thesis.Then,the training set is smoothed with Savitzky-Golay,so that the Gaussian process regression is not affected by the output noise of the encoder-decoder model,and the capacity decay trend can still be predicted.Finally,the predicted values of the two models are weighted summed according to a set of dynamic weights.The experimental results show that the fusion prediction method can effectively reduce the cumulative error.