Transfer learning applied to DRL-Based heat pump control to leverage microgrid energy efficiency
Paulo Lissa, Michael Schukat, Marcus Keane, Enda Barrett
Abstract
Domestic hot water accounts for approximately 15% of the total residential energy consumption in Europe, and most of this usage happens during specific periods of the day, resulting in undesirable peak loads. The increase in energy production from renewables adds additional complexity in energy balancing. Machine learning techniques for heat pump control have demonstrated efficacy in this regard. However, reducing the amount of time and data required to train effective policies can be challenging. This paper investigates the application of transfer learning applied to a deep reinforcement learning-based heat pump control to leverage energy efficiency in a microgrid. First, we propose an algorithm for domestic hot water temperature control and PV self-consumption optimisation. Secondly, we perform transfer learning to speed-up the convergence process. The experiments were deployed in a simulated environment using real data from two residential demand response projects. The results show that the proposed algorithm achieved up to 10% of savings after transfer learning was applied, also contributing to load-shifting. Moreover, the learning time to train near-optimal control policies was reduced by more than a factor of 5.