Test Trajectory Optimization for Parameterizing a Neural Network-Based Equivalent Circuit Battery Model

This paper introduces a test trajectory optimization approach for the parameterization of a machine learning-enhanced lithium-ion battery model. The model embeds a dense neural network within an equivalent circuit model (ECM) of battery dynamics. This hybridization strikes a balance between the interpretability of the ECM versus the ability of the dense neural network to provide a detailed nonlinear representation of open-circuit voltage (OCV) versus state of charge (SoC). The study's main contribution is optimizing battery cycling to reduce the extensive data requirements for hybrid model parameterization. Towards this goal, the paper performs Fisher analysis on all the model parameters simultaneously. Poor parameter identifiability is observed, particularly due to the inherent redundancy of neural network parameters. A novel optimization method is then employed to minimize the additional information needed for a well-conditioned Fisher information matrix, as opposed to directly maximizing Fisher information. The effectiveness of the proposed approach is validated through simulation, demonstrating its potential to enhance model parameterization significantly. Copyright (c) 2024 The Authors.