Skip to main content
U.S. Department of Energy
Office of Scientific and Technical Information

Neural network adaptive controllers for energy management

Conference · · Proceedings of the American Power Conference; (United States)
OSTI ID:5067953
 [1]
  1. Honeywell Sensor and System Development Center, Minneapolis, MN (US)
Energy management of buildings and plants requires controllers that can perform regulatory control of heating or cooling systems whose dynamic behavior may be complex and may vary with the nature of each plant or installation. In general, a controller that can adapt to the behavior of each plant would be necessary to optimize energy savings. The control problem to be solved is the regulatory control of an unknown plant model with the goal of maximizing comfort and energy savings. In this paper, the authors present a class of adaptive controllers, based on neural networks that use reinforcement learning, that are self-tuning since they do not require knowledge of the desired control function. In conjunction with an unsophisticated monitor, this neural-network-based controller can develop near-optimal control strategies on-line by exploration of the control space without assumptions about the plant model. The authors show how such controllers can be used for regulatory control and for control scheduling in continuous processes. Experiments in control of an environmental controller according to a desired time-temperature schedule demonstrate good learning performance. The authors propose how these controllers can be designed for solving a broader class of energy management problems.
OSTI ID:
5067953
Report Number(s):
CONF-9104106--
Conference Information:
Journal Name: Proceedings of the American Power Conference; (United States) Journal Volume: 53
Country of Publication:
United States
Language:
English