Intra-day Bidding Strategies for Storage Devices Using Deep Reinforcement Learning

The problem faced by the operator of a storage device participating in a continuous intra-day (CID) market is addressed in this paper. The goal of the storage device operator is the maximization of the cumulative rewards received over the entire trading horizon, while taking into account operational constraints. The energy trading is modeled as a Partially Observable Markov Decision Process. An equivalent state representation and high-level actions are proposed in order to tackle the variable number of the existing orders in the order book. The problem is solved using deep reinforcement learning (RL). Preliminary results indicate that the agent converges to a policy that scores higher total revenues than the ``rolling intrinsic''.