Automatic Load Management in Active Distribution Grids Using Reinforcement Learning

Demand side management (DSM) is a key component for integrating volatile and intermittent renewable energy sources in power grids. In this paper, we propose and study DSM in the form of indirect load control where an agent automatically controls household devices based on real-time electricity prices. The major challenge is the limited available information that includes only past and current values of the household’s net consumption and the electricity prices. To handle this issue, DSM is performed using a reinforcement learning-based approach. The results show that the decision agent performs close to optimal for input data similar to the training data. For input data that diverges from the training data, satisfactory performance is achieved when changing the building parameters, but not when the price curves are shifted. An alternative training approach is suggested that solves this issue.