Reinforcement Learning for Building Energy System Control in Multi-Family Buildings
Abstract
Replacing the widespread use of fossil fuel heating systems in the existing multi-family building stock with systems powered by renewable energy is one of the most important contributions to the energy transition in Germany. Heat pumps can also be a suitable solution for older buildings, especially in combination with photovoltaic systems. Advanced control strategies are needed to optimize the self-consumption rate and reduce the load on the electricity grid. Due to the increase in available computing power, artificial intelligence methods such as reinforcement learning, which can be used to control (energy) systems, have become very popular in the recent years. A literature review was conducted to provide an up-to-date overview of the use of reinforcement learning for building energy system control, with a focus on multi-family buildings. Such a building with an energy system including a heat pump, thermal and electrical storage, and a photovoltaic system was modeled in MATLAB/Simulink. The model is used to investigate typical control strategies as a reference and to develop a heat pump control strategy using the MATLAB Reinforcement Learning Toolbox. Two reinforcement learning agents with different numbers of observations were trained to control the heat pump and studied in annual simulations. Both RL agents provide the required supply water temperature. Adding one observation improves the annual reward by 160 %. Nevertheless, a potential to improve learning the RL agent switching of the heat pump at times of high bottom storage temperatures was identified.
Keywords
Multi-family building, Energy system, Optimization, Self-consumption, Reinforcement learning