On-site estimation of battery electrochemical parameters with physics-informed neural networks in dynamic current profiles

Abstract

The accurate on-site estimation of battery electrochemical parameters is crucial for optimal battery management, enabling advanced control strategies and reliable prognostics. However, physics-based methods often suffer from high computational costs, require specific testing setups; while data driven solutions lack interpretability, creating a need for solutions including the benefits of both strategies [1]. We present a novel framework for on-site physical parameter estimation, for real-time characterization of lithium-ion batteries, leveraging on the recent attention for hybrid physics-based and data-driven solutions. Our approach utilizes a two-phase modelling strategy that combines Physics-Informed Neural Networks (PINNs) with transfer learning [2]. In an initial ”data-agnostic” phase, a PINN is trained exclusively using the governing physical equations of a single particle model. The model is set to include Fourier Feature transformations on the dependent variables, so that we extend the learning range to dynamic current profiles. During the second phase, critical ageing-related electrochemical parameters are fine-tuned using real-world voltage profile data. This two-phase strategy significantly reduces computational cost compared to traditional optimization methods, making it suitable for implementation on Battery Management Systems, and the dataagnosticism of the initial training phase avoids the need for large chunks of data. We demonstrated the framework’s efficacy through the estimation of diffusivities and active material volume fractions. Experimental and analytical validations showed a relative accuracy of 3.89% in estimating the active material volume fractions. Furthermore, our proposed PINN-based approach outperformed classical optimization techniques in accurately recovering parameters under varied ageing conditions.