Operational optimization for 3E-efficient desiccant air-conditioning systems based on deep reinforcement learning

Decoupling humidity control from sensible cooling in office buildings offers significant potential for energy savings, but the complex dynamics of desiccant air-conditioning systems (DACS) challenge conventional control methods. This study develops a control-oriented deep reinforcement learning framework for optimizing a DACS coupled with a combined cooling, heating, and power (CCHP) plant, with DSAC-T adopted as the learning core. The framework aims to improve energy, economic, and environmental performance while maintaining indoor thermal comfort. To enable stable long-horizon training with complex physical models, the framework is implemented through a robust Python-Modelica co-simulation environment with a two-level fault-tolerance scheme proposed in this study. Operating within a shared supervisory control logic to ensure consistent mode switching and fair comparison, the agent learns to dynamically modulate chilled water and regeneration air temperature setpoints in response to time-varying operating conditions. Comparative analysis against a conservative fixed-setpoint baseline reveals that the intelligent strategy reduces primary energy consumption by 11.4%, operating costs by 14.4%, and carbon emissions by 12.7%. Furthermore, the DSAC-T strategy demonstrates superior stability compared to a rule-based strategy, especially under high-load conditions. Supplementary perturbation tests and a benchmark comparison against standard SAC, DDPG, and TD3 further support its practical robustness and algorithmic suitability. These findings show that the proposed DSAC-T strategy can effectively coordinate the strongly coupled thermodynamic processes of the studied DACS-CCHP system, thereby alleviating the conflict between stringent indoor comfort standards and system-wide multidimensional efficiency.