Collaborative optimization and regulation of the passenger cabin thermal environment by integrating machine learning modelling and the KMIGA control strategy

Computational Fluid Dynamics (CFD) serves as the mainstream method for developing thermal comfort in passenger cabins. However, its lengthy computational cycles and low simulation efficiency pose significant challenges in coupling with thermal management systems, thereby restricting the rapid iteration and engineering application of thermal management control strategies. To tackle this issue, this paper proposes an innovative reduced-order modelling approach based on machine learning for predicting thermal comfort in passenger cabins. Taking an SUV model as a case study, a multi-condition sample space is constructed using Latin hypercube sampling. A dual neural network-based Predicted Mean Vote (PMV) prediction model is then designed and trained, exhibiting strong adaptability and high prediction accuracy under varying environmental temperatures, solar radiation intensities, and air supply parameters. To extend the applicability across different vehicle models, this paper introduces a model generalization scheme. By inputting a small amount of CFD simulation data from a new vehicle, the source prediction model can be rapidly migrated and optimized, enabling quick adaptation to the new vehicle's passenger cabin structure and meeting the prediction demands for multi-vehicle thermal comfort. Finally, this paper proposes a KMIGA (Kriging-Multi-Island Genetic Algorithm) collaborative optimization control strategy, which effectively balances air conditioning energy consumption and passenger thermal comfort. Validation results show that during the early stages of control strategy development, the machine learning reduced-order model can replace the CFD model for efficient optimization. In real-world driving tests under WLTC conditions, the proposed strategy reduces energy consumption by 4.9% compared to PID control, confirming its practical viability.