Neural Battery for Energy Storage System Modeling Based on Hidden-State Dynamic Process Solver

Abstract

The development of precise models for simulating rapidly expanding systems has become imperative for enhancing the planning and utilization of energy storage. It is often the case that traditional physical models are not suitable for use in calculations involving large or complex battery systems. This work proposes a neural battery model, which is developed by constructing a battery hidden-state dynamic process solver based on a neural network. The model overcomes the explicit dependence of conventional physics-driven approaches on model assumptions and governing equations. Instead, it employs a latent state space to uniformly characterize the internal dynamics. The implementation of dynamic process solving frameworks, such as neural ordinary differential equations (Neural ODEs), facilitates the establishment of a hidden-state dynamic system that ensures numerical stability and accuracy. Moreover, a battery network computational framework is proposed, which utilizes parallel computing to overcome the efficiency limitations of the model for large-scale battery packs.