NeuroECM: A Physics-Informed Neural Network Fusion Model for Equivalent Circuit Model of Li-Ion Battery

The paper presents NeuroECM, an equivalent circuit model (ECM) for Li-ion batteries that integrates a Long Short-Term Memory (LSTM) network with Multi-Layer Perceptron (MLP) networks to enhance battery state estimation. The LSTM processes historical current sequences to estimate State-of-charge (SOC), while the MLP networks leverage automatic differentiation to learn battery model parameters. Unlike traditional Mean Squared Error (MSE) based optimization, a weighted smooth Mean Absolute Error (wMAE) loss function is introduced, which prioritizes early-stage accuracy to mitigate long-term error accumulation. Additionally, we enforce physical constraints through a boundary loss to ensure realistic voltage predictions. The proposed NeuroECM framework is validated using 18650 Li-ion battery experimental data, demonstrating highly accurate SOC estimation with an MAE of 0.03963 and terminal voltage prediction with an MAE of 0.04592. The results show that NeuroECM offers high accuracy while ensuring interpretability, making it a promising methodology for realtime battery management.