An Event-Driven Neural Kalman Model for State Representation and Learning-Based Dynamic Scheduling of Industrial Energy System

Operation optimization of industrial energy systems (IES) can effectively improve energy utilization efficiency and reduce carbon emissions. Learning-based optimization models such as reinforcement learning (RL) have been widely applied in dynamic scheduling of energy systems. However, they are usually time-driven periodic optimization that require the state variables having obvious time distribution rules, without considering the event-triggered characteristics of IES, and the uncertainty in state distribution caused by production-energy coupling. Thus, an event-driven neural Kalman state representation and dual-scale strategy learning framework is proposed in this study. To implement latent state construction with time distribution characteristics, an improved proximal policy optimization with neural Kalman state representation is developed to formulate time-driven scheduling policies, where the Kalman filter and RL parameters are trained in cascade to achieve an end-to-end learning process. In order to realize an event-driven policy learning, a state transition matrix connecting approach is proposed for collaborative calculation of time-event dual-scale policies. To verify the effectiveness of the proposed method, real data from a domestic steel company are employed for experiments. Commonly used and state-of-the-art approaches are selected for comparisons. The results validate the advantages of the proposed approach in terms of scheduling frequency, cost-effectiveness, and adaptation to industrial uncertain environments.