An Intelligent Learning Mechanism for Trading Strategies for Local Energy Distribution

Abstract

Although most electric power is presently generated using fossil fuels, two abundant renewable and clean energy sources, solar and wind, are increasingly cost-competitive and offer the potential of decentralized (and hence more robust) sourcing. However, the intermittent nature of solar and wind power presents difficulties in connection with integrating them into national power grids. One approach to addressing these challenges is through an agent-based architecture for coordinating locally-connected energy micro-grids, each of which manages its own local energy production, distribution, and storage. By integrating these micro-grids into a larger network structure, there is the opportunity for them to be more responsive to local needs and hence more cost effective overall. In such an arrangement, the micro-grids have agents that can choose to resell their excess energy in an open, regional market in alignment with respect to their specific goals (which could be to reduce carbon emissions or to maximize their financial outcomes). In this study, we have investigated how agents operating in such an open environment can learn to optimize their individual trading strategies by employing Markov-Decision-Process-based reasoning and reinforcement learning. We empirically show that our learning trading strategies improve net profit loss by up to 29 % and can reduce carbon emissions by 78 % when compared to the original (non-learning) trading strategies.