Use of optimised MLP neural networks for spatiotemporal estimation of indoor environmental conditions of existing buildings

Abstract

Controlling the indoor environmental quality in real time is essential for the health, well-being and productivity of occupants of a building. In recent years, research has focused on improving monitoring devices and strategies and developing techniques for estimating indoor conditions. The use of machine learning algorithms in this context has increased considerably. However, monitoring data in real time from large multizone working areas is challenging. The aim of this work is to provide an interpolation methodology based on the use of optimised multilayered perceptron neural networks to estimate the indoor environmental conditions of a building in real time. These estimations are obtained without the need for neither monitoring in the occupied working area nor human intervention and considering low-cost sensors. The neural network is optimised by implementing the multiobjective genetic algorithm NSGA-II to find the best architecture in terms of error and complexity. This method was applied to the building of a research centre in north-western Spain, where interpolated values for indoor air temperature, relative humidity and CO2 concentration were obtained. The results of this case study yielded relative errors close to 6% for temperature, 5% for relative humidity, and 12% for CO2 concentration. These values validate the methodology developed for the estimation of indoor environmental conditions and the contribution of this research to the improvement of the monitoring and control of the indoor environmental quality of a building.