Two-stage scheduling of integrated energy systems based on a two-step DCGAN-based scenario prediction approach

Abstract

Integrated energy systems (IESs) are developing rapidly as a supporting technology for achieving carbon reduction targets. Accurate IES predictions can facilitate better scheduling strategies. Recently, a newly developed unsupervised machine learning tool, known as Generative Adversarial Networks (GAN), has been used to predict renewable energy outputs and various types of loads for its advantage in that no prior assumptions about data distribution are required. However, the structure of the traditional GAN leads to the problem of uncontrollable generations, which can be improved in deep convolutional GAN (DCGAN). We propose a two-step prediction approach that takes DCGAN to achieve higher accuracy generation results and uses a K-means clustering algorithm to achieve scenario reduction. In terms of scheduling strategies, common two-stage scheduling is generally day-ahead and intraday stages, with rolling scheduling used for the intraday stage. To account for the impacts on the prediction accuracy of scheduling results, Conditional Value at Risk (CVaR) is added to the day-ahead stage. The intra-day prediction process has also been improved to ensure that the inputs for each prediction domain are updated in real-time. The simulations on a typical IES show that the proposed two-step scenario prediction approach can better describe the load-side demands and renewable energy outputs with significantly reduced computational complexity and that the proposed two-stage scheduling strategy can improve the accuracy and economy of the IES scheduling results.