SmartCorners - Report on validated simulation models (D1.4)

This deliverable aims to capture the trade-off between cabin comfort (cooling performance) and energy-efficiency using a smart control algorithm based on artificial intelligence (AI). A causal-based AI is trained within a simulation environment, utilizing validation methodologies from Tasks T1.3 and T1.4, to optimally control the thermal system actuators with minimal operational cost. To accelerate computational performance, reduced-order fast running models (FRMs) of the thermal system—derived by AVL DE from high-fidelity models developed in Task T1.3—are integrated into the simulation framework. The optimisation of the thermal management control system is approached in a stepwise manner through the application of reinforcement learning (RL). An initial subsystem of the overall thermal management control architecture is selected for substitution by an RL-based control agent. This focused implementation serves as proof of concept, allowing the RL evaluation in a controlled and limited scope before broader integration. By applying RL to a specific control domain within the thermal system, insights are gained into performance improvements, stability, and learning dynamics. This stepwise approach ensures a systematic and scalable transition from rule-based to intelligent control, minimizing system-level risks while leveraging the potential of RL for predictive and energy-efficient operation. A RL-based framework is under development that aims to combine thermal simulation environments with state-of-the-art RL-based optimisation algorithms. RL agents are designed to optimise control strategies in the form of machine learning (ML) neural networks. By combining these powerful RL libraries with SmartCorners’ thermal simulation environments, this deliverable provides solutions for Assets 2 & 4 in the SmartCorners project and therefore achieves objective O4. The thermal plant models incorporate user-specific passenger comfort models which can be tailored to each new passenger individually. The trained agent can incorporate passenger comfort as one of several targets and therefore optimise control functions that are as individual as the passenger digital twin. For the training of optimal control policies, surrogate models are used instead of the actual high-fidelity simulation environment. This is particularly interesting if the simulation model is complex, and simulation time is close to real-time. Particularly in early development stages, it is beneficial to use fast surrogate models and generate fast approximations. Multiple ML models have been used to train surrogate models for control targets and for optimisation (substitution to the slow physical models). The implemented RL framework has been designed in a way that both high-fidelity simulation models and ML-based surrogate models could be used to train the same agents, thereby optimally enabling the transition between different development stages.

Hali tarjima qilinmagan