A hybrid learning strategy for the control of window blinds

Abstract

Automated control of building systems is gaining increasing attention due to rising energy costs and enhanced environmental consciousness. Automated window blinds have the potential to save energy through striking a balance between daylighting levels and air conditioning load. Currently, most control systems for window blinds operate by setting a limit on the maximum direct solar radiation coming through the window (typically 100W/m 2) or on estimated glare discomfort. Trade-off between the energy required for artificial lighting and air conditioning is not currently evaluated through a global optimization framework. This paper presents a methodology for the control of window blinds using global optimization involving lighting and energy simulations making use of minimum number of sensors. Since these simulations take a lot of time, it is not recommended for real-time control. Instead, machine learning techniques are used to speed up the computations of optimal control actions. A combination of hierarchical clustering and perceptron networks is used to develop a generalized representation of potential solutions which are evaluated off-line. A case-study of an office building is used to evaluate the advantages of the approach. It is shown that significant improvement in efficiency is possible. The average prediction error was less than 15 for the selected case study. © 2009 ASCE.