The smart grid enables flexible demand and controlling devices based on fluctuating prices, resulting in savings for individuals and reduced carbon emissions. A challenge is optimising the control of devices, requiring predicting future prices and demand, and optimising decisions over a longer time horizon.

This project will build on existing work and focus on the engagement with the user in smart grid autonomous systems. The challenges here are three-fold. First of all, for the user to trust the decisions that are made by the autonomous system and engage with the system in a constructive way, these decisions need to be presented in an intelligible and easy to understand manner. Techniques from explainable artificial intelligence will be used for this purpose.

Second, to produce optimal decisions on behalf of the user, it needs to understand the user's preferences and constraints, e.g. when they are at home, what their comfort level is, and trade offs between costs and convenience. The system can infer these preferences indirectly by observing the user, or it can ask explicitly, but this causes additional burden to the user. This part of the project will explore preference elicitation techniques combined with human-in-the-loop reinforcement learning to infer the user model with minimal input from the user.

Finally, the project will consider the user incentives to reduce carbon emissions. The project will use techniques from game theory to ensure that the user has no incentive to manipulate the system (e.g. by misrepresenting their flexibility for demand or personal comfort preferences).