Forecasting Indoor Air Quality in Mexico City Using Deep Learning Architectures

Air pollution causes millions of premature deaths per year due to its strong association with several diseases and respiratory afflictions. Consequently, air quality monitoring and forecasting systems have been deployed in large urban areas. However, those systems forecast outdoor air quality while people living in relatively large cities spend most of their time indoors. Hence, this work proposes an indoor air quality forecasting system, which was trained with data from Mexico City, and that is supported by deep learning architectures. The novelty of our work is that we forecast an indoor air quality index, taking into account seasonal data for multiple horizons in terms of minutes; whereas related work mostly focuses on forecasting concentration levels of pollutants for a single and relatively large forecasting horizon, using data from a short period of time. To find the best forecasting model, we conducted extensive experimentation involving 133 deep learning models. The deep learning architectures explored were multilayer perceptrons, long short-term memory neural networks, 1-dimension convolutional neural networks, and hybrid architectures, from which LSTM rose as the best-performing architecture. The models were trained using (i) outdoor air pollution data, (ii) publicly available weather data, and (iii) data collected from an indoor air quality sensor that was installed in a house located in a central neighborhood of Mexico City for 17 months. Our empirical results show that deep learning models can forecast an indoor air quality index based on outdoor concentration levels of pollutants in conjunction with indoor and outdoor meteorological variables. In addition, our findings show that the proposed method performs with a mean squared error of 0.0179 and a mean absolute error of 0.1038. We also noticed that 5 months of historical data are enough for accurate training of the forecast models, and that shallow models with around 50,000 parameters have enough predicting power for this task.