Physics-Informed Meta-Reinforcement Learning for Adaptive Thermal Management of Battery Energy Storage Systems

Abstract

We propose a physics-informed meta-reinforcement learning framework for adaptive thermal management of battery energy storage systems. Conventional battery thermal management systems rely on fixed rule-based or proportional-integral-derivative controllers that cannot adapt to evolving cell degradation or highly variable operating conditions. The proposed method replaces these static controllers with an autonomous agent that continuously learns and adjusts its cooling strategy. Our architecture integrates a differentiable physics-informed neural network as a surrogate model of the coupled electro-thermal dynamics within the battery pack. This surrogate is trained on high-fidelity simulation data and fine-tuned with streaming sensor measurements, thereby enforcing physical consistency through the governing heat conduction equation. A gated recurrent unit module further encodes slow temporal drifts in cell parameters—such as increasing internal resistance and decreasing heat capacity due to aging—into a latent state representation. The policy network then receives this augmented state and outputs continuous commands for coolant pump speed and phase-change material activation threshold. We train the entire system using a model-agnostic meta-learning procedure across a curriculum of synthetic aging distributions. This meta-training enables zero-shot transfer of thermal control strategies to previously unseen degradation states without requiring additional retraining. The reward function penalizes excessive cell temperatures, thermal gradients, and unnecessary pumping power, thereby balancing safety, uniformity, and energy efficiency. Our approach maintains the existing battery thermal management hardware while fundamentally restructuring the control logic. The system runs at one hertz on an edge artificial intelligence processor co-located with the battery management system controller. This work contributes a self-calibrating, degradation-aware thermal management solution that can generalize across diverse usage scenarios and aging trajectories, potentially extending battery lifetime and improving operational safety in large-scale energy storage installations.