Learning-Enhanced Nonlinear Model Predictive Control for Battery Thermal Management Systems

Abstract

Battery thermal management (BTM) systems in electric vehicles are required to regulate the temperature of the battery powering the vehicle. Model predictive control (MPC) is an optimization-based control strategy that has proven useful in nonlinear control tasks across many different domains, and is therefore a promising candidate for BTM. However, battery thermal management systems are difficult to model due to nonlinearities, and simplified control models that do not fully capture the true dynamics are often employed, which can result in reduced control performance. In this thesis, an adaptive control framework is proposed for learning model residuals using a neural network. The learned residuals are used within the control model of the controller, resulting in a control model that adapts to the system. Specifically, the neural network is trained using two distinct loss functions, resulting in two distinct adaptive controllers. Both adaptive controllers are compared against a nominal controller relying solely on a physics-based model, on both matched and mismatched systems. The framework is initially tested on a benchmark reference tracking cascaded tank system, where it successfully learns the mismatch in dynamics and achieves improved closed-loop control performance. The framework is subsequently evaluated for both reference tracking and economic MPC formulations in BTM systems. For reference tracking, the adaptive controllers yielded mixed results, in some scenarios decreasing cost by up to 44 %, whereas in other scenarios increasing cost by up to 409 %. Similarly the root-mean-squared tracking error was reduced in some cases, and substantially increased in others. In economic MPC, the adaptive controller achieved cost reductions of 23 % to 35 % for all mismatched models, while incurring up to 11 % higher cost in a scenario with a matched model. Model adaptation via neural network residuals is therefore not automatically beneficial, as the approach is sensitive to the loss design, hyperparameters, and the training data. The proposed framework does improve performance in some scenarios, and enhancing its robustness and generalizability warrants further investigation.