Predicting spatiotemporal concentrations in a multizonal residential apartment using conventional and Physics-informed deep learning approach

Most indoor air pollution studies focusing on modeling and material balance assume well-mixed conditions, which is usually not true in larger and multizonal spaces. Spatially nonhomogenous concentrations can lead to considerably different personal exposure of occupants within the same indoor space. Studying the interzonal transport of pollutants and their governing factors provides critical insights into the fate and transport of pollutants. The current work focuses on predicting PM2.5 and CO2 concentrations in different zones of a residential apartment using measured concentrations in one zone using conventional and physics-informed long short-term memory (PI-LSTM) models for different internal door configurations. Model predictions were validated using experimentally obtained spatiotemporal data sets using the exposure and maximum concentration (relative to measured) as key performance metrics. The PI-LSTM model performed better in most cases for PM2.5, while the LSTM model exhibited better predictive accuracy for CO2 concentrations. As more internal doors were opened and the number of zones increased, PI-LSTM’s predictive accuracy declined. PM2.5 predictions were more accurate for zones near the emission source than those farther away.