Data-driven modelling of battery state-of-health using multi-criteria-based feature reduction

Previous studies have explored many features that can be used to estimate the health of lithium batteries. However, there are still gaps from previous research, namely, which features should be used and which can be ignored to make the best SOH estimation using machine learning models. This paper proposes a multi-criteria-based feature reduction method to find and combine the best features with machine learning models for estimating the lithium ion battery SOH. This research consists of three main stages: first, determining the features that will be used for building the model, these features include voltage, current, temperature, and time; secondly, carrying out multi-criteria-based feature reduction, existing features are selected based on a combination of four methods such as Pearson correlation rank, Lasso regression, sequential feature selection (SFS), and PCA; third, using the selected features to test the battery health estimation performance using multi-layer perceptrons. The results show that the proposed multi-criteria-based feature reduction method can determine useful features, thereby increasing the generalisation ability and accurate prediction results for lithium-ion battery health degradation under actual EV usage conditions. Besides, the proposed method combined with MLP can outperform other models to 40% of R2.