State-of-charge estimation of batteries using parameterized physics-informed neural networks

Accurate state-of-charge (SOC) estimation is essential for battery management systems (BMS). Conventional model-based methods suffer from initialization sensitivity and parameter dependency, while data-driven methods require extensive datasets and lack physical interpretability. This paper proposes a hybrid neural network method that integrates temporal sequence learning with parameterized physics-informed neural networks (PPINN) to estimate the SOC of the public LG 18650HG2 and laboratory-tested INR 21700-33 J lithium-ion batteries. The proposed method employs a hypernetwork-based framework that generates dynamic weights by jointly encoding SOC-dependent baseline parameter sets and LSTM-based temporal predictions, enabling adaptive learning across diverse operating conditions. This combined encoding allows a latent representation of battery states to effectively control the solution network, while physical dynamics are directly incorporated into the loss function to ensure consistency. Experimental results across various temperature conditions and driving cycles show superior performance, achieving RMSE values between 0.83% and 1.59% for the full public dataset. Furthermore, the proposed method exhibits exceptional robustness to missing data, achieving an RMSE as low as 0.92% even after a 62% reduction of the public training set. Notably, the PPINN maintains a robust RMSE of less than 2.51% in laboratory-tested scenarios, even under extreme data-scarce conditions. • PPINN framework resolves parameter dependency in battery state estimation • Hypernetwork generates adaptive weights for diverse operating conditions • Physical governing equations are integrated directly into the loss function • Latent manifold representation captures generalized battery operating scenarios • Physics-based constraints ensure robust learning with limited training data