A Novel Auto-LSTM-Based State of Health Estimation Method for Lithium-Ion Batteries

Abstract

Abstract Lithium-ion batteries (LIBs) have been widely applied in modern society. The state of health (SOH) estimation can provide helpful guidance to maintain LIBs in advance. Machine learning (ML) and deep learning (DL) have been widely applied to pursue the high accuracy SOH estimation. However, the accuracy and performance of ML/DL methods heavily rely on their hyperparameters, and the hyperparameters tuning process for ML-/DL-based SOH estimation is mainly optimized by manual search, which are very time consuming and can hardly find the good hyperparameters configuration within the limited time resource. In this study, a new automatic long short-term memory (LSTM) method, called auto-LSTM, is developed for the SOH estimation, which can tune the hyperparameters in feature selection, LSTM structure, and its training algorithm in the automatic way. First, a LSTM model is developed for the SOH estimation. Second, the hyperparameters of the proposed LSTM are collected to be optimized by random search (RS) and tree Pazen estimator (TPE) automatically. Third, as the hyperparameters of auto-LSTM are characteristic as the hierarchy high dimension, a novel hyperparameter reduction algorithm (HRA) is developed to promote RS and TPE. The proposed auto-LSTM is tested on the NASA dataset and CALCE dataset. The results show that the proposed auto-LSTM with HRA can promote both RS and TPE on most case studies, validating its potential for providing a user-friendly and easy method for the SOH estimation on LIBs.