Decision tree and Parametrized classifier for Estimating occupancy in energy management

Abstract

A new kind of supervised learning approach is proposed to determine the number of occupants in a room in order to use these estimate for improved energy management. It introduces the concept of Parametrized classifier. It relies on the predetermined structure of supervised learning classifiers, where any classifier could be used to evaluate this approach. The parameters will be adjusted according to the incoming data sensors (i.e CO2 concentration, acoustic pressure, …) using a tuning mechanism depends on an optimization process. This paper provides different supervised learning methods (i.e decision tree random forest) to determine the required structure in order to be used in parametrized classifier approach. The structure of decision tree has been chosen which represents the classification rules and limit the depth of the tree to facilitate the generalization process. In order to evaluate the generalization possibilities of a supervised learning approach (i.e. decision tree), it has been chosen to extrapolate results from office H358 to another similar office H355. The knowledge has been extracted from a decision tree built on H358 office then applied and tuned for H355 using parameterized classifier approach. Moreover, experiments implement occupancy estimations and hot water productions control show that energy efficiency can be increased by about 6% over known optimal control techniques and more than 26% over rule-based control besides maintaining the occupant comfort standards. The building efficiency gain is strongly connected with the occupancy estimation accuracy.