2015/04/08 by Frederik Ruelens, Ruelens, Frederik, Bert Claessens +9
Energy · Engineering · #Energy Efficiency and Management #FOS: Computer and information sciences #FOS: Electrical engineering #Machine Learning (cs.LG) #Refrigeration and Air Conditioning Technologies #Smart Grid Energy Management #Systems and Control (eess.SY) #Traffic control and management #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.1504.02125
openalex publication_date 2015/04/08 · openalex created_date 2022/10/02 · openalex updated_date 2026/07/28
Driven by recent advances in batch Reinforcement Learning (RL), this paper\ncontributes to the application of batch RL to demand response. In contrast to\nconventional model-based approaches, batch RL techniques do not require a\nsystem identification step, which makes them more suitable for a large-scale\nimplementation. This paper extends fitted Q-iteration, a standard batch RL\ntechnique, to the situation where a forecast of the exogenous data is provided.\nIn general, batch RL techniques do not rely on expert knowledge on the system\ndynamics or the solution. However, if some expert knowledge is provided, it can\nbe incorporated by using our novel policy adjustment method. Finally, we tackle\nthe challenge of finding an open-loop schedule required to participate in the\nday-ahead market. We propose a model-free Monte-Carlo estimator method that\nuses a metric to construct artificial trajectories and we illustrate this\nmethod by finding the day-ahead schedule of a heat-pump thermostat. Our\nexperiments show that batch RL techniques provide a valuable alternative to\nmodel-based controllers and that they can be used to construct both closed-loop\nand open-loop policies.\n