State of health and remaining useful life estimation of lithium-ion battery based on parallel deep learning methods

Abstract

Accurate estimation of state of health (SOH) and remaining useful life (RUL) is the key to efficient and safe utilization of lithium-ion batteries. Neural networks are widely used to estimate SOH and RUL of lithium-ion batteries due to their strong nonlinear fitting, adaptive and self-learning abilities. Convolutional neural networks (CNN) could effectively extract the key features and recurrent neural networks (RNN) have excellent ability to capture time information in series data, but the traditional series models combining CNN and RNN may cause information leakage in time series data due to the lack of time association learning ability in CNN. In this paper, we propose a novel parallel architecture integrating temporal convolutional networks (TCN) with long short-term memory (LSTM) to address the temporal information leakage limitation in conventional series models. This synergistic design enables simultaneous extraction of spatial degradation features through TCN and dynamic temporal dependencies via LSTM, significantly enhancing SOH and RUL estimation accuracy for lithium-ion batteries compared to sequential CNN-RNN approaches. The comparison of SOH and RUL estimation performances between series and parallel models is conducted using public battery datasets sourced from the National Aeronautics and Space Administration (NASA) and Oxford university. The improvement of SOH estimation accuracy across the entire dataset is more than 29 % in dataset 1 and 37 % in dataset 2. For RUL estimation, the parallel models could achieve comparable performance to series models, with an acceptable increase in time consumption.