Interpretable AI for explaining and predicting battery state of health using PSO-enhanced deep learning models

Abstract

Accurately determining the State-of-Health (SOH) of Lithium-ion (Li-ion) batteries is important for the safe operation of Electric Vehicles (EVs); nevertheless, in practical implementations, variables such as human error and operational conditions can affect the accuracy of SOH estimates. This paper proposes a fusion neural network based on the Convolutional Neural Network, Long-term Short Memory (LSTM), and Convolutional LSTM (ConvLSTM) models with meta-heuristic optimization and eXplainable Artificial Intelligence (XAI) to accurately predict the battery SOH. Our proposed model combines CNN, LSTM, and ConvLSTM in a novel process, optimized by PSO and explained by SHAP, in contrast to previous hybrid models that often only combine two neural networks and rarely incorporate both optimization and interpretability. The suggested method integrates CNN, LSTM, and ConvLSTM models into a Deep Neural Network (DNN) framework. This framework is tuned using Particle Swarm Optimization (PSO) to improve the generality and accuracy of SOH estimations. Furthermore, in order to achieve serial data time dependency and correlation, the suggested data-driven approach exploits the voltage distribution and capacity changes in the derived battery discharge curve. The results of the experiment show that when compared to CNN, LSTM, ConvLSTM, and other deep learning models, the proposed model achieves good performance for battery SOH prediction. Furthermore, the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) are limited to within 0.009% and 0.044%, accordingly, on the battery dataset. This study utilizes XAI techniques, namely SHapley Additive exPlanations (SHAP), to clarify the predictions made by the fusion DNN model. This approach aims to enhance clarity and instill trust in the system, and the findings indicate that the fusion DNN model outperforms conventional approaches, optimized using PSO and including CNN, LSTM, and ConvLSTM components, and it achieves higher $R^2$ scores, smaller mean residuals, and improved XAI outcomes.