Dynamic Energy Management in Competing Microgrids using Reinforcement Learning

• Main Authors: Vivek, V P, Diddigi, Raghuram Bharadwaj, Bhatnagar, Shalabh
• Format: Conference Proceeding
• Language: English
• Subjects: Batteries; dynamic pricing; finite horizon; Heuristic algorithms; Microgrids; Reinforcement learning; Renewable energy sources; smart grid; Smart grids; stochastic games; Wind power generation
• ISSN: 2472-8152
• DOI: 10.1109/ISGT59692.2024.10454198

Smart Grids are a concept for efficiently managing energy in power grids using both renewable and non-renewable power sources. Integrating renewable power sources into the traditional power supply has multi-fold advantages, like reducing carbon emissions and minimizing the peak demand at the main grid. An efficient way for this integration is by employing microgrids, which are autonomous intelligent decision-making networks that cater power to a small subset of customers. The microgrids are equipped with a renewable source like solar or wind to generate power and battery sources to store power. The microgrids would need to make optimal decisions like the amount of power to be stored in the battery, selling price, etc., at each decision time instant to maximize their long-term goal of making a profit. At the same time, from the customers' perspective, fairness is paramount for easier and faster adoption. In this work, we address both problems in a unified framework by proposing a novel smart grid paradigm where two microgrids compete to distribute power to their customers. We formulate the competition in a two-player zero-sum stochastic game framework and propose an efficient Reinforcement Learning algorithm for solving this problem. Moreover, our formulation and solutions are designed for the finite horizon rather than the infinite horizon, thereby ensuring practical applicability. We demonstrate the performance of the algorithm through extensive experiments on simulated smart grid settings.