Fusing Machine Learning Methods with Computational Fluid Dynamics and Sensor Data for Indoor Air Quality Monitoring

Monitoring indoor air quality is important nowdays, as noted by many researchers, since pollution potentially introduces a major impact on the health of people. There are many factors that increase the concentration of a pollutant in a room and there may even arise an additive effect if many of them co-exist. Thus, monitoring the presence and the dispersion of a pollutant in a room may reveal causes of the increment of the pollutant and the exposure of people. These may be used to find effective solutions for pollution mitigation. There are two main methodologies for performing the analysis of this problem: methods that rely strictly on mathematical computations based on laws of physics, also know as Computational Fluid Dynamics (CFD) simulations, which potentially suffer from high computational cost, and data-driven methods that rely on the redundant availability of detailed pollution data, while not addressing Fluid Dynamics laws. However, a whole spectrum of methodologies lies between these two extremes that can effectively reduce the negative effects of the two extremes, while combining the positive ones. In this paper we investigate the combination of CFD simulation data and Machine Learning algorithms in real-world scenarios using sensors for estimating the dispersion of pollutants in rooms. We present an application for a such a scenario, where Machine Learning models have been trained on CFD data and they are used for a given classroom along with a sensorial system to get real time estimation of the concentrations in the whole room. As shown by the real world experiments we have conducted, our hybrid approach achieves small estimation times compared to raw CFD calculations, whereas the Mean Squared Error of the calculations is quite low at the pre-development stage.