Nonlinear Fractional Dynamics Integrated Physics-Informed Neural Network Model for LiFePO4 Batteries in Electric Vehicles

This paper addresses the long-standing challenge of attaining high-precision models for LiFePO4 Batteries which suffer from weakly observable dynamics. We introduce a new paradigm of integrating a nonlinear fractional-order physics-based model with a hybrid neural network model. First, a fractional-order model (FOM) is proposed to capture the physics of the battery that existing integer-order models (IOMs) fail to replicate, such as the solid phase diffusion. The FOM parameters are state dependent as they vary along with the progression of the state of charge (SOC). Second, the unknown and unmodelled physics is captured by a hybrid neural network model integrated with the FOM. The physical states of the FOM are used to guide the neural network resulting in a state dependent nonlinear fractional-order physics-informed neural network (FO-PINN) to predict the terminal voltage of the battery. Validation with experimental results and comparisons with existing modelling techniques reveal that the proposed scheme delivers improved predictive accuracy with decreased computational cost and enhanced physically meaningful information. The scheme has potential in applications that demand high propulsive power and accuracy, such as electric aircraft.