Neural Network-Based Imitation Learning for Approximating Stochastic Battery Management Systems

Abstract

Lithium-ion batteries play a pivotal role in enabling eco-friendly mobility, particularly in electric vehicles, but optimizing their charging process to improve battery lifespan, safety, and overall efficiency remains a significant challenge. Traditional predictive control methods are limited by their reliance on precise models, which are often hindered by uncertainties in battery parameters due to aging, production variability, and operational conditions. While stochastic predictive control policies can address these uncertainties by incorporating them directly into the optimization process, they typically introduce considerable computational complexity. In response to this challenge, this paper presents a novel approach that adapts imitation learning to efficiently approximate stochastic predictive control strategies, thus significantly reducing the computational burden through offline training. Specifically, the proposed method leverages the Dataset Aggregation algorithm to overcome the issue of distributional shift, a common limitation in imitation learning frameworks. Simulations based on a detailed electrochemical model demonstrate the effectiveness of the method, adhering to probabilistic constraints while offering a scalable and computationally efficient solution for advanced battery management systems.