Reinforcement Learning-Based Hyperparameter Tuning for Adaptive Model Predictive Controllers in Battery Thermal Management

Battery thermal management (BTM) is crucial for maximizing the battery performance and maintaining the battery health. Model Predictive Control (MPC) stands as a promising technique for achieving this objective. However, conventional MPC approaches employing fixed parameters often restrict the control performance. This paper aims to adaptively tune the MPC parameters for an optimal BTM system using reinforcement learning. To achieve this, the nonlinear MPC controller is linearized at nominal operating points and transformed into an adaptive MPC controller. Leveraging a soft actor-critic reinforcement learning agent, the MPC decisive parameters, including weights of the cost function, lengths of the prediction horizon and control horizon, are regulated online. Simulation results demonstrate that the proposed scheme achieves a 5.8% reduction in energy consumption for BTM, an 18.8% decrease in constraints violation, and a substantial 59% reduction in MPC execution time compared to an MPC controller with fixed parameters. Additionally, the feasibility of the proposed reinforcement learning-driven adaptive MPC has been verified on a dSPACE machine. These outcomes underscore the significant potential of MPC optimization for enhancing battery management systems.