Hybrid Physics–Informed and Machine Learning Model for Accurate Lithium‐Ion Battery Voltage Prediction

Exact modeling of lithium‐ion batteries is essential for the optimal design and functioning of contemporary energy storage systems. This research introduces a hybrid modeling approach that integrates an extended Shepherd equivalent circuit model (ECM) with a multilayer perceptron (MLP) neural network to improve voltage prediction precision. The ECM parameters are determined utilizing the Red‐Tailed Hawk (RTH) optimization algorithm, a contemporary metaheuristic that exhibits enhanced convergence efficacy relative to conventional methods. The MLP is designed to rectify residual voltage prediction errors by accounting for nonlinearities and dynamic phenomena overlooked by the physical model. The suggested hybrid approach is evaluated employing experimental data from a commercial Enertech SPB58253172P2 lithium‐ion battery (3.75 V, 20 Ah) under dynamic current profiles. An ablation study is performed to demonstrate the impact of network depth on the accuracy, with the (256, 128, 64) architecture providing the best performance. Also, the performance will be assessed against the decision tree and random forest algorithms. The results indicate a substantial decrease in prediction error, with the root mean square error (RMSE) declining from 0.1521 V to 66.6 mV and the mean absolute error (MAE) reducing from 0.1373 V to 53.4 mV. These findings underscore the model's potential for incorporation into sophisticated battery management systems (BMS), especially under dynamic operating situations.