Industrial Energy Management and Production Decision Making via Lyapunov-Guided Learning

Energy-intensive industries have to reduce fossil fuel consumption while scheduling production for cost efficiency. It poses the question that how to coordinate renewable energy generation, storage, heat recovery and energy cascade utilization in real time to deal with the low energy efficiency and continuous production problems existing in complex dynamic coupled process production. This question is further complicated while facing difficulties in collaborative modeling and online control by the underlying stochastic process without accurate statistic knowledge. To characterize the above issues, a non-convex operation optimization problem is formulated for coupling production and energy joint scheduling. To obtain a simple online solution with provable performance, a method by combining Lyapunov optimization and actor-critic deep reinforcement learning is proposed. The former is used to decouple the original problem into small-size non-convex subproblems for each time slot and guarantee the long-term constraints. The latter aims at the non-convex part by using model information of the former to obtain accurate evaluations of production actions for fast convergence and high robustness with low computational complexity. The simulation shows that the proposed method can achieve the online optimal benefit while ensuring production tasks and system stability with high scalability.