State of health estimation of lithium-ion batteries based on multi-feature extraction and temporal convolutional network

Abstract

Accurate estimation of the state of health (SOH) is vital for maintaining the safe and stable operation of lithium-ion batteries. However, the utilization of extant data-driven methods for SOH estimation often poses a dichotomy between the diversity of feature selection and the intricacy of network models. This study proposes an estimation methodology combining multi-feature extraction with temporal convolutional network (TCN). The experimental curves of charging and discharging, alongside the incremental capacity curve of lithium-ion batteries, were subjected to principal component analysis to extract Class I features. Class II features were derived by performing empirical mode decomposition on the battery capacity decay curve, thereby securing multi-feature data. Moreover, a channel attention module based on TCN was utilized to process multi-dimensional features and select appropriate weights. Concurrently, to enhance the adaptive threshold training ability of the model with multiple input parameters, a residual shrinkage network was introduced. The SOH estimation of lithium-ion batteries was ascertained by training and processing these multi-features using an improved TCN. The results were subsequently compared with long short-term memory and conventional TCN models. The proposed model demonstrated a mean absolute percentage error of 1.47 % in estimating the SOH of lithium-ion batteries.