Multi-Pollutant Air Quality Prediction Using a ResNet-Based Regression Model: Evaluation Using Abuja and Beijing Air-Quality Datasets

Abstract

- Prediction of air quality is fundamental to the study of pollution dynamics as well as the implementation of monitoring strategies. In this paper, the construction and performance of a One-dimensional Residual Neural Network (1D ResNet) based multi-output regression model that predicts multiple air pollutants using historical air quality data are reported. Two multi-site datasets with distinct spatial and temporal characteristics have been used to assess the performance of the model, namely, Abuja, Nigeria, and Beijing, China. For the Abuja case study, the model predicted the concentrations of PM1, PM10, and PM2.5 with a location aware and station balanced model approach, while the Beijing case study predicted the levels of PM2.5, PM10, and SO2. Both case studies independent test sets for final evaluation. The Abuja model achieved an overall R² of 0.6774, with pollutant-specific R2 values of 0.6721, 0.6885, and 0.6715 for PM1, PM10, and PM2.5, respectively. For Beijing, R2 values of 0.8939, 0.8567, and 0.7117 were achieved for PM2.5, PM10, and SO2, respectively. It is evident from the results that the level of prediction success depends on the type of pollutants, the location of the monitoring stations, and the data situation, where there was a decrease in the prediction success levels during extreme pollution levels.