Evaluation of artificial intelligence-enhanced pid tuning for virtual thermal system calibrationSkobir, Tilen (Avtor) Ramšak, Matjaž (Mentor) Semenič, Tilen (Komentor) reinforcement learningPID controlvehicle thermal managementelectrified vehicleco-simulationFMUTD3reward shapingThis master's thesis presents the development and evaluation of an artificial intelligence methodology for autonomous PID controller calibration in a multi-loop vehicle thermal management system (VTMS). A co-simulation environment was constructed linking an electrified vehicle thermal model, a software control model, and a reinforcement learning agent with TD3 algorithm. The agent autonomously searched for optimal values of four PI gains for both the electronic expansion valve (EXV) and the compressor. The reward function was developed iteratively through four configurations. Compared to the initial engineer's gain approximation, the best episode reduced the mean absolute error (MAE) by 54%. The results confirm that a properly designed RL agent can autonomously find stable multi-loop PID gains, and that constructing a reliable co-simulation environment represents the primary technical bottleneck of the workflow.20262026-08-20 16:04:43Magistrsko delo/naloga99588sl