A constraint-oriented, reward-enhanced reinforcement learning framework for real-time energy management in continuous industrial processes

Deep reinforcement learning (DRL) has strong potential for real-time demand response (DR) in industrial energy management, but its practical deployment is often limited by slow and unstable training. To address this challenge, this work presents a constraint-oriented, reward-enhanced (CORE) framework and instantiates it with deep deterministic policy gradient (CORE-DDPG). The proposed approach enforces production constraints and reduces reliance on manual reward engineering by introducing a dual-level expert-guided reward structure derived from a quadratic programming (QP) model. In a real-world case study, CORE-DDPG consistently outperformed representative DRL baselines and QP-based benchmarks. Compared with the without-DR baseline, CORE-DDPG achieved an average daily saving of $599.12 (2.70%), outperforming the day-ahead QP's $476.35 (2.02%) and the rolling QP's $505.37 (2.16%) under real prices. These improvements were statistically significant as confirmed by one-tailed paired t-tests (p = 0.035 and p = 0.048, respectively). CORE-DDPG also maintained robust performance against intra-day temporal sequence perturbations in electricity prices. Overall, the proposed approach provides a state-driven and economical solution for real-time control in continuous industrial processes.