Fast Multi-Period Security-Constrained Economic Dispatch Based on Deep Neural Networks

Abstract

Security-constrained economic dispatch (SCED) is one of the most important daily tasks for operators. The scale of security constraints is huge for practically-sized power systems, which makes the SCED difficult or even impossible to be solved. Whereby, the number of active security constraints is relatively small. By eliminating the inactive security constraints, the complexity of SCED can be significantly reduced. Focusing on it, this paper proposes an intelligent framework to accelerate the calculation of SCED without any loss of accuracy. The proposed framework uses a deep neural network (DNN) to reduce the online computational cost significantly by shifting the heavy computation into offline training. More specifically, a DNN is used to extract the feature of SCED, which can effectively pre-identify the active constraints of SCED model. Moreover, an efficient lightweight learning strategy is presented to improve the learning efficiency of the DNN by the feature selection and feature decomposition. The effectiveness of the proposed method is demonstrated in modified IEEE benchmark systems.