Prediction of school PM by an attention-based deep learning approach informed with data from nearby air quality monitoring stations.
Journal:
Chemosphere
PMID:
39999669
Abstract
Predicting indoor air pollutants concentrations in schools is essential for ensuring a healthy learning environment. Traditional measurements methods pose challenges in cost, maintenance, and time. This study proposes a new approach using a deep learning (DL)-based soft sensor to predict PM concentrations in school environment both indoor (classroom) and outdoor (playground). The proposed soft sensor, based on attention deep convolutional autoencoder (ADCAE), leverages data from nearby monitoring stations, completely eliminating the need for school on-site sensors and the attendant financial and technical costs of installation, operation, and maintenance. Results indicate superior predictive performance of the ADCAE model compared to traditional machine learning methods, with significantly higher coefficient of determination (R), lower mean squared error and lower mean absolute error across all studied schools. The prediction accuracy reached 0.97, 0.976, and 0.936 for indoor PM in the elementary school, middle school, and high school, respectively, while the school outdoor PM showed even a higher prediction performance with R values of 0.9923, 0.9869, and 0.943. Thus, the developed soft sensor based on DL is a promising tool to effectively monitor air quality in schools and to capture response measures to fine dust pollution at schools.