Transfer learning for Demand Response of a Multi-Agent Battery and Electric Water Heater System

Renewable Energy Sources (RES) are ever more finding their way into today's power systems. They have many benefits, but hold challenges when it comes to operating the power system. Their intermittency defies the current prominent approach of keeping system balance; supply follows demand. Demand Response (DR) has been proposed numerous times to mitigate the challenges of RES. Additionally, Reinforcement Learning (RL) has been proposed numerous times to mitigate the scalability challenges of DR. Simultaneously, RL has been criticised for its data inefficiency. In an effort to tackle this problem, we showcase a transfer learning approach in a DR setting. The application consists of a household, equipped with inflexible load, a rooftop solar installation, an Electric Water Heater (EWH) and a battery. The household's electrical energy consumption is billed according to a Time-of-Use-pricing scheme. In this multi-agent DR application, two RL-agents, one for the EWH and one for the battery, have to efficiently operate the system. We show that in our scenario RL, combined with pretraining of the battery-agent, reduces operation cost by 28.5 %, compared to rule-based control.