Reinforcement learning-based intelligent control algorithm for energy-efficient and emission-aware optimization in industrial power systems
Abstract
This paper proposes a novel Reinforcement Learning–Based Intelligent Control (RL-IC) algorithm for energy-efficient and emission-aware optimization in industrial power systems. Unlike conventional rule-based or model-predictive controllers that rely on fixed dynamics, RL-IC employs a hierarchical dual-objective learning architecture that jointly optimizes fuel efficiency and emission stability. The lower-level controller performs real-time actuation through a deterministic policy, while the high-level coordinator adjusts emission constraints via adaptive reward shaping. An attention-based interpretability layer enhances transparency, enabling human operators to trace decision rationale and improve safety assurance. Experimental evaluation on a 660 MW digital twin power unit demonstrates that RL-IC achieves 9.8% lower fuel consumption and 98.4% compliance within emission targets compared with advanced DRL baselines such as PPO, SAC, and DDPG. Furthermore, RL-IC exhibits faster convergence, higher robustness under dynamic load fluctuations, and transferable performance across heterogeneous plants. The results confirm that hierarchical reinforcement learning, when combined with interpretability and safety constraints, provides an effective paradigm for sustainable and adaptive industrial automation.