Safety-constrained minimization of CHP auxiliary power consumption ratio via high-fidelity surrogate-assisted deep reinforcement learning

Reducing the Auxiliary Power Consumption Ratio (APCR) is critical for enhancing the energy efficiency of Combined Heat and Power (CHP) units. However, the practical deployment of intelligent control strategies remains hindered by the stochastic noise inherent in industrial sensor data and the zero-tolerance of physical systems for trial-and-error exploration. To address these challenges, this study proposes a safety-guaranteed Model-based Deep Reinforcement Learning (MBRL) framework. First, to mitigate non-stationary data disturbances, a high-fidelity surrogate model integrating RevIN and BiLSTM-Attention is constructed as a virtual interaction environment. This model achieves a coefficient of determination (R2) of 0.9702 on noisy test datasets, establishing a trustworthy foundation for agent training. Second, to guarantee operational safety, an embedded hard-constraint safety projection layer is introduced. By forcibly projecting actions onto the physically feasible domain, this mechanism ensures strict physical compliance within the action space. Finally, a TD3-based multi-variable cooperative optimization strategy is designed. By leveraging gradient sensitivity analysis to pinpoint critical control dimensions and employing a survival pressure-based asymmetric threshold reward, this strategy effectively overcomes policy stagnation. Experiments using real-world data from a 350 MW CHP unit demonstrate that the proposed method reduces the APCR by an average of 0.1980% while maintaining operational safety, demonstrating clear superiority over the classic DDPG algorithm (0.0598%). This reduction translates into substantial annual operational cost savings, fully validating the economic viability and safety of the proposed paradigm for industrial applications.