The Adaptive Setback Thermostat
We present an adaptive setback thermostat (AST) which switches between two temperature setpoints — one optimized for user comfort and one for saving energy. The AST operates locally in an office room and makes its decisions based on how the room is (expected to be) used. Core issues, in decreasing order of importance, are user comfort, user friendliness (ease of installation and use) and to reduce energy costs. It is argued why a reinforcement learning approach may not be the best solution, and then shown how to reformulate the problem using a simple heuristic where reward maximization is replaced by explicit prediction of user arrivals, using temporal difference learning.
Unable to display preview. Download preview PDF.
- Mozer MC, Vidmar L, Dodier RH. The Neurothermostat: Predictive optimal control of residential heating systems. In: Mozer MC, Jordan MI, Petsche T (Eds), Advances in Neural Information Processing Systems 9, MIT Press, 1997, pp 953–959Google Scholar
- Sutton RS, Barto AG. Reinforcement Learning: An Introduction, MIT Press, 1998.Google Scholar
- Sutton RS. Learning to Predict by the Methods of Temporal Difference, Machine Learning 1988; 3:9–44, Kluwer Academic PublishersGoogle Scholar
- Lögdahl P. The Adaptive Setback Thermostat: Experiments in simulated and real office environments. MSc Thesis, Dept. of Computer Systems, Uppsala University, Sweden, 1998Google Scholar