Multi-objective deep reinforcement learning for a water heating system with solar energy and heat recovery

Abstract

Deep reinforcement learning (DRL) has gained attention from the scientific community due to its potential for optimizing complex control schemes. This study describes the implementation of a DRL platform that allows training smart agents to manage a complex water heating system in an institutional building that uses solar energy and waste heat from a water chiller as energy sources. In addition to optimizing the use of energy while delivering hot water, the agents are also trained to activate the chiller, to avoid the premature degradation of the system, and to continue providing hot water in case of failures of heating devices of the system. The definition of the reward function, which is fundamental to simultaneously impose all these goals on the agent, is tested by comparing different agents trained to prioritize different goals. In comparison with a previous study done on the same system with standard reinforcement learning techniques, this method allows far more freedom to control the system. The deep neural networks and the DRL algorithm were programmed without specialized libraries, implying that the algorithm could be used to train smart agents in programs without direct access to deep learning libraries, or in the actual system with a simple programmable controller.