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

Sammanfattning

Battery thermal management (BTM) systems in electric vehicles are required to regulatethe temperature of the battery powering the vehicle. Model predictive control(MPC) is an optimization-based control strategy that has proven useful in nonlinearcontrol tasks across many different domains, and is therefore a promising candidatefor BTM. However, battery thermal management systems are difficult to model dueto nonlinearities, and simplified control models that do not fully capture the truedynamics are often employed, which can result in reduced control performance.In this thesis, an adaptive control framework is proposed for learning model residualsusing a neural network. The learned residuals are used within the control modelof the controller, resulting in a control model that adapts to the system. Specifically,the neural network is trained using two distinct loss functions, resulting intwo distinct adaptive controllers. Both adaptive controllers are compared againsta nominal controller relying solely on a physics-based model, on both matched andmismatched systems. The framework is initially tested on a benchmark referencetracking cascaded tank system, where it successfully learns the mismatch in dynamicsand achieves improved closed-loop control performance. The framework issubsequently evaluated for both reference tracking and economic MPC formulationsin BTM systems.For reference tracking, the adaptive controllers yielded mixed results, in some scenariosdecreasing cost by up to 44 %, whereas in other scenarios increasing cost byup to 409 %. Similarly the root-mean-squared tracking error was reduced in somecases, and substantially increased in others. In economic MPC, the adaptive controllerachieved cost reductions of 23 % to 35 % for all mismatched models, whileincurring 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 trainingdata. The proposed framework does improve performance in some scenarios, andenhancing its robustness and generalizability warrants further investigation.