Adaptive State of Charge and State of Health Estimation through Optimized Deep Reinforcement Learning with Dynamic Agent Selection and Test-Time Training Tailored to Dynamic Operating Conditions

Abstract Accurately estimating the state of charge (SOC) and state of health (SOH) of lithium−ion batteries under dynamic conditions is critical for electric vehicles (EVs) and energy storage systems; however, nonlinearity and uncertainties hinder traditional methods. This paper proposes an adaptive estimation framework using optimized deep reinforcement learning (DRL). Key innovations include integrating large language models (LLMs) with test-time training (T3), a process of continuous online model adaptation, for cross-chemistry generalization, model-free DRL with dynamic agent selection and data sharing, solutions for high distribution shifts in aged batteries, self-supervised health index (HI) learning via attention mechanisms, and a hybrid DRL architecture with dual SOC/SOH networks. The algorithm combines pre-trained DRL, T3, reward shaping, and experience replay. Extensive experiments on six battery chemistries under varying temperatures (−10−50 °C), noise (0.3−1%), and aging scenarios demonstrate the framework’s superiority, achieving a state of charge (SOC) error of ≤1.34% and a state of health (SOH) mean absolute error (MAE) of <4%, even for batteries with >30% capacity fade. Case studies validate the adaptability to real-world challenges. While hardware integration remains, this advances real-time battery management for safety and efficiency.