An Extended Single-Particle Model Based on Physics-Informed Neural Network for SOC State Estimation of Lithium-Ion Batteries

Because of its great accuracy, the electrochemical model is frequently utilized in the algorithm design process for lithium-ion batteries. Sadly, the electrochemical model requires a lot of time to solve since it is made up of many nonlinear partial differential equations. In order to solve an extended single particle model (ESPM) fast, a neural network based on physical information (PINN) is examined in this paper. The PINN-ESPM structure can not only estimate the state of charge, but also quickly and accurately estimate the lithium-ion concentration and potential under various application currents, which has stronger adaptability and scalability. In the process of neural network learning, different from the traditional neural network that needs to be trained by labeled data, the loss function is designed only based on the physical constraints brought by equations, boundary conditions and initial values, which makes it an unsupervised learning method. Finally, by comparing the PINN-ESPM proposed in this paper with the data obtained by the P2D model under various current conditions and the experiment battery voltage, the maximum relative error is maintained at 4%. The error of SOC based on the model is less than 4%. While under the same computing resources, PINN-ESPM is 500 times faster than the traditional numerical method.