Title:A Holistic Power Optimization Approach for Microgrid Control Based on Deep Reinforcement Learning

Abstract:The global energy landscape is undergoing a transformation towards decarbonization, sustainability, and cost-efficiency. In this transition, microgrid systems integrated with renewable energy sources (RES) and energy storage systems (ESS) have emerged as a crucial component. However, optimizing the operational control of such an integrated energy system lacks a holistic view of multiple environmental, infrastructural and economic considerations, not to mention the need to factor in the uncertainties from both the supply and demand. This paper presents a holistic datadriven power optimization approach based on deep reinforcement learning (DRL) for microgrid control considering the multiple needs of decarbonization, sustainability and cost-efficiency. First, two data-driven control schemes, namely the prediction-based (PB) and prediction-free (PF) schemes, are devised to formulate the control problem within a Markov decision process (MDP). Second, a multivariate objective (reward) function is designed to account for the market profits, carbon emissions, peak load, and battery degradation of the microgrid system. Third, we develop a Double Dueling Deep Q Network (D3QN) architecture to optimize the power flows for real-time energy management and determine charging/discharging strategies of ESS. Finally, extensive simulations are conducted to demonstrate the effectiveness and superiority of the proposed approach through a comparative analysis. The results and analysis also suggest the respective circumstances for using the two control schemes in practical implementations with uncertainties.