Demand Response for Flexibility Carriers using Reinforcement Learning

Abstract

In recent work concerning demand response for flexibility carriers, specialized methods have achieved considerable progress for specialized instances of carriers. These methods often rely on complicated design decisions in feature engineer- ing or utility function design. Furthermore, flexibility carriers are often stochastic in their behaviour. In this paper we propose a general model-free reinforcement learning approach using limited feature engineering and a straightforward utility function. We validate our approach on a simulation of a flexibility carrying cold storage cell. Our results indicate significant cost savings can be achieved through our approach, at the cost of a long exploration period. Our approach requires ap- proximately 69 simulated days before offering an improvement in cost over stan- dard carrier behaviour.