An Optimal Scheduling Framework for Integrated Energy Systems Using Deep Reinforcement Learning and Deep Learning Prediction Models
Lei Zhang, Ye He, Hongbin Wu, Nikos Hatziargyriou
Abstract
Model-free deep reinforcement learning has emerged as a promising method for addressing the scheduling challenges in integrated energy systems. However, uncertainty in system states continues to hinder optimization efforts. This paper proposes a hybrid framework that integrates deep learning prediction models with deep reinforcement learning scheduling models. Initially, Gaussian process regression is employed to extract interval information from stochastic variables, organizing the input data into structured time series segments. Subsequently, a hybrid prediction model, combining Transformer and long short-term memory networks, is constructed for multi-step interval prediction of these stochastic variables. Finally, a synchronized training mechanism couples the twin delayed deep deterministic policy gradient method with the hybrid prediction model, fully exploiting trends in system state changes to enhance scheduling performance. Experiments are conducted to validate the effectiveness of the proposed framework using open-source data to analyze the influence of different prediction methods on scheduling. Results show that the proposed approach improves overall performance by 22.4% compared to the baseline method