Artificial neural networks for predicting air flow in a naturally ventilated test room

Abstract

The objective of this research is to investigate air flow distribution inside a light weight test room which is naturally ventilated using artificial neural networks. The test room is situated in a relatively sheltered location and is ventilated through adjustable louvres. Indoor air temperature and velocity are measured at four locations and six different levels. The outside local temperature, relative humidity, wind velocity and direction are also monitored. The collected data are used to predict the air flow across the test room. A multi-layer feedforward neural network was employed with three hidden slabs. Satisfactory results with correlation coefficients equal to 0.985 and 0.897, for the indoor temperature and combined velocity, respectively have been obtained when unknown input data, not used for network training, were used as input. Both values are satisfactory especially if the fact that combined velocity readings were very unstable is considered. The work presented in this paper primarily aims to show the suitability of neural networks to perform such predictions. In order to make the method more usable the training database needs to be enriched with readings from actual measurements from a number of applications.