Physics-Informed Multi-Head Attention-Based SOC Estimation for LFP Batteries Using Only Measured Signals

Accurate state-of-charge (SOC) estimation for LiFePO4(LFP) batteries remains challenging due to the flat open-circuit voltage (OCV)-SOC plateau and hysteresis, which reduce SOC observability from terminal voltage measurements. This paper proposes a multi-head-attention physics-informed neural network (MHA-PINN) for SOC estimation that relies only on measurable battery-management-system signals (current, terminal voltage, and temperature) and does not require prior knowledge of the initial SOC or internal concentration as inputs, supporting practical onboard deployment. The estimator combines an attention-based sequence model with a physics residual derived from the Coulomb-counting SOC dynamics, thereby enforcing physically consistent SOC trajectories while improving robustness against drift and overfitting. The approach is validated on an LFP BYD 4680 cell (15 Ah) using a dataset spanning multiple drive cycles and operating modes. Across unseen test cases, including the Urban Dynamometer Driving Schedule (UDDS), Federal Urban Driving Schedule (FUDS), Highway Fuel Economy Test (HWFET), constant-current (CC) charge/discharge, and constant-power (CP) operation, the proposed method achieves RMSE values below 2% (0.00656-0.01990) and consistently outperforms both a purely data-driven attention baseline and a Gated Recurrent Unit (GRU)-based physics-informed reference model.