Learning-Aided Framework for Storage Control Facing Renewable Energy

Abstract

The Internet of Things (IoT) enables reliable and fast data collection and transmission, providing key infrastructure for power generation, distribution, and control in the smart grid. This IoT-enabled smart grid tackles challenges brought by renewable penetration in new ways: Accurate and real-time information allows for the application of artificial-intelligence-powered computation. We employ the deep learning framework and consider the problem of storage control facing uncertainties in renewable generation. We propose both model-based and model-free storage control frameworks to identify the value of information. For the first framework, opposing to most deep-learning-oriented research in the electricity sector, we use the one-shot load decomposition technique to encode structural information into the learning framework. The structural information refers to the fact that the one-shot load decomposition maintains the control strategy space. Based on this structural information, we develop the storage control policy by utilizing a deep learning framework for price and renewable prediction, which is the basis of our deep-learning-enabled storage control. For the model-free framework, we regard historical price and demand data as input and directly output the control actions. For each model, we further establish theoretical analysis on how the uncertainties in price and renewables influence the cost. Numerical evaluations illustrate the remarkable performance of our proposed frameworks and reveal the value of information.