CNN-based state prediction for a varying number of storage in economic dispatch

Economic dispatch (ED) is essential for power system operations. However, the large-scale energy storage (ES) integration introduces numerous binary state variables into ED formulations. Although relaxation-based methods and machine learning techniques have been developed to alleviate the computational burden from ES binary variables, the former is restricted due to critical application conditions that may not hold in practice, and the latter cannot deal with a varying number of ES in the real-world deregulation of electricity markets. To this end, this paper proposes a data-driven state prediction method for a varying number of ES in an ED problem. A symmetrical convolutional neural network (CNN) structure is leveraged to learn the states from the concatenation of the input load and operating limits. Such an architecture effectively extracts multi-scale features through multiple convolutional layers, pooling, and upsampling operations. It excels at handling the coupled relationships between input loads and states with information from other time intervals. A projection layer is further designed to adjust the CNN outputs to handle a varying number of ES. The effectiveness of the proposed method is demonstrated in the IEEE 118-bus test system and a real-world 661-bus utility system. • A data-driven method for state prediction with varying ES counts in ED is proposed. • A symmetrical convolutional neural network structure is leveraged to predict the states. • A projection layer is designed to adjust the outputs for varying ES counts. • A better accuracy-efficiency trade-off than KNN/RNN baselines is demonstrated.