Deep Reinforcement Learning for Scenario-Based Robust Economic Dispatch Strategy in Internet of Energy

Abstract

Currently, the integration of distributed energy generators through virtual power plants in the Internet of Energy is a mainstream method. The complex structure of virtual power plants and the characteristics of distributed energy make it difficult to solve the economic dispatch problems of virtual power plants. In addition, the load of a virtual power plant is unstable and uncertain and thus requires a robust economic dispatch strategy. Because the selection of the set of uncertain conditions is conservative, the traditional robust economic dispatch strategies cannot effectively reduce the cost of virtual power plants. In addition, the traditional methods for solving robust strategies cannot directly solve nonlinear and nonconvex problems. In this article, we propose a scenario-based robust economic dispatch strategy for virtual power plants, aiming to reduce the operational costs of virtual power plants. First, to reduce the conservatism of the strategy, scenario-based data augmentation is adopted for data generation. Through a generative adversarial network, a large amount of scene data are generated to extend the set of uncertain conditions. The scene data cannot only reduce the conservatism but also can be used in the determination of robust strategies. Second, deep reinforcement learning is adopted for historical data training, directly solving nonlinear and nonconvex problems to obtain a robust economic dispatch strategy. As experiments show, with the accurate generation of scene data, the proposed economic dispatch strategy is robust and effectively reduces the cost of virtual power plants.