Multi-agent residential demand response based on load forecasting

Improving the efficiency of the smart grid, and in particular efficient integration of energy from renewable sources, is the key to sustainability of electricity provision. In order to optimize energy usage, efficient demand response mechanisms are needed to shift energy usage to periods of low demand, or to periods of high availability of renewable energy. In this paper we propose a multi-agent approach that uses load forecasting for residential demand response. Electrical devices in a household are controlled by reinforcement learning agents which, using the information on current electricity load and load prediction for the next 24 hours, learn how to meet their electricity needs while ensuring that the overall demand stays within the available transformer limits. Simulations are performed in a small neighbourhood consisting of 9 homes each with an agent-controlled electric vehicle. Performance of agents with 24-hour load prediction is compared to the performance of those with current load information only and those which do not have any load information.