Treffer: Joint optimization for temperature and humidity independent control system based on multi-agent reinforcement learning with cooperative mechanisms.

The coupling of indoor thermal environment parameters and the complexity of subsystem control pose challenges to further enhancing the energy efficiency of air conditioning systems. This study adopts a multi-agent deep reinforcement learning algorithm for optimal control of a temperature and humidity independent control system combining radiant ceiling cooling with desiccant dehumidification. In the co-simulation process, Python completes the algorithm construction and employs the system and building models constructed by EnergyPlus to achieve the training of agents. This study aims to analyse the optimisation effects of different algorithms, discuss the behavioural characteristics among multiple agents during the training process, and reveal the impact of humidity intelligent control on the energy saving and energy-use flexibility effects. Compared to conventional single-agent optimization control methods, this study integrates interdependent control actions and leverages a cooperative mechanism to enhance system optimization. Simulation results demonstrate that the method can avoid the local cognitive limitations of individual agents in dynamic environments, achieving multi-objective optimization. Notably, the humidity control agent autonomously reduces humidity to assist in lowering the supply water temperature, thereby managing high sensible heat loads while avoiding condensation risks of the radiant cooling panels. The optimized strategy reduces chiller energy consumption by 41.4%, with 16% of this reduction attributed to humidity control. Furthermore, humidity control enables the chiller to operate at varying power levels, enhancing the energy-use flexibility. This study provides a replicable and transferable solution for multi-subsystem optimal control. • Optimization of temperature and humidity independent control systems is performed. • The robustness of multi-agent deep deterministic policy gradient is proved. • Collaborative mechanism allows the agent to reduce the chiller energy use by 44%. • Humidity control can enhance the effect of energy saving and control flexibility. [ABSTRACT FROM AUTHOR]