A Novel Approach for Optimization of Smart Micro-grid Using Multi Agent Reinforcement Learning

In this paper we consider grid connected solar micro-grid system which contains a consumer, a solar photovoltaic system and a battery. The consumer is considered as an agent who continuously interacts with the environment and learns to take best actions. Initially Multi Agent System is implemented for effective energy management of solar micro-grid. Then reinforcement learning is imparted to the agent to make it smart. Each agent uses a model-free reinforcement learning algorithm, namely Q Learning, to optimize the battery scheduling in dynamic environment of load and available solar power. Multiple agents sense the states of the environment components and make coordinated, collective decisions about how to respond to randomness in load, intermittent solar power, unexpected events and unplanned actions using a Multi-Agent Reinforcement Learning algorithm, called Coordinated Q Learning (CQL). The goal of each agent is to increase the utility of the battery and solar power in order to reduce the power consumption from grid. Each agent individually optimizes and contributes to global optimization. Simulation results using real numerical data are presented for improving the reliability and stability of solar micro-grid under dynamic environment. Also substantial reduction in the grid power requirement is proved leading to economic and environmental optimization.