Remaining Useful Life Prediction Of Lithium-ion Batteries Based On Gated Recurrent Neural Network

Rechargeable lithium-ion batteries are widely used in portable electronic devices,electric vehicles,and other fields because of their high energy density,long service life,and other advantages.However,after long-term use of lithium-ion batteries,due to a variety of internal electrochemical reactions,the physical properties will be changed,resulting in battery performance degradation and equipment failure.The remaining useful life prediction of lithium-ion battery can effectively evaluate its future working ability and provide decision guideline for predictive maintenance of potential faults.Therefore,it is of great practical significance to accurately predict the remaining useful life of lithium-ion batteries for fault warning and safe operation of battery-related equipment and systems.Among the existing literature methods,model-based remaining useful life prediction methods are limited by the complexity of mechanism modeling,which makes it difficult to obtain accurate degradation models.However,the data-driven method does not need to comprehend the complex mechanism of battery degradation,and uses monitoring data to mine the degradation information and realize the remaining useful life prediction.In this paper,based on the data-driven approach with the battery capacity as the dominant variable,the remaining useful life prediction research of lithium-ion batteries is conducted based on the gated recurrent unit of gated recurrent neural network as the elementary prediction model.The main research contents are as follows:(1)To address the problems of non-smooth,non-linear,multi-modal and multi-noise characteristics in the remaining useful life prediction of lithium-ion batteries that affect the prediction accuracy,a method of remaining useful life prediction of lithium-ion batteries based on decomposition strategy and hybrid neural network is proposed.Firstly,multi-scale decomposition of the battery capacity data is used to obtain different component sequences by using variational mode decomposition.Then the capacity component with the highest correlation is screened out as modeling features using Pearson correlation coefficients.Finally,a novel hybrid neural network model based on gated recurrent unit is designed to predict the remaining useful life of lithium-ion batteries with the capacity component within the time window as the input and the capacity component at the next moment of the time window as the output.The experimental results show that the proposed method can reduce the influence of the above complex features on the prediction,and effectively improve the accuracy of the remaining useful life prediction of the batteries.(2)To address the problems of unreliable selection of prediction sliding window size and neglected quantification of prediction uncertainty in the remaining useful life prediction of lithium-ion batteries,a method of remaining useful life prediction of lithium-ion batteries based on data sliding window with Monte Carlo dropout is proposed.Firstly,the sliding window size of the battery capacity is calculated using Cao’s method to construct the capacity data sliding window.Then the dropout method is used in the aforementioned hybrid neural network and the model is trained with the capacity data in the sliding window as the input and the next capacity of the sliding window as the output.Finally,in the prediction stage,the hybrid neural network model incorporating the dropout method is used to perform multiple Monte Carlo simulations to quantitatively calculate the mean and variance of the prediction results,and then obtain the 95% confidence interval and probability distribution for the prediction of the remaining useful life of batteries to achieve the quantification of the prediction uncertainty.The experimental results show that the proposed method can not only effectively describe the uncertainty in battery remaining useful life prediction,but also the prediction results are more accurate with less error.(3)To address the problems that the capacity is difficult to be measured directly and the importance of different features is not considered in indirect prediction in the remaining useful life prediction of lithium-ion batteries,a method of remaining useful life prediction of lithium-ion batteries based on multi-source features and dual-stage attention is proposed.Firstly,six health indicators characterizing battery aging are extracted from the charge and discharge voltage,current,temperature and time data of the battery as multi-source features.Then an encoder-decoder network model incorporating dual-stage attention mechanism and gated recurrent unit is developed for remaining useful life prediction of lithium-ion batteries with the above multi-source features as input and capacity as output.The experimental results show that the proposed method can effectively replace the capacity to achieve multi-source feature prediction,adaptively weigh the feature importance,further improve the accuracy of battery remaining useful life prediction,and has certain interpretability.