Forecasting the air quality time series data using univariate and multivariate convolutional neural network models

This study compares three different deep learning approaches for forecasting the air quality time series data. All the approaches were based on the convolutional neural network (CNN), which is one of the deep learning algorithms. The first approach, known as the univariate CNN (UCNN), is based solely on the time series past values. Meanwhile, the second approach, known as the multi-parallel CNN (MPCNN), will use the same variable time series from different observations as the input for the model. The last approach is the multivariate CNN (MCNN), where the forecast will incorporate other variables' time series as the input for the model. The time series data used in this study are the hourly particulate matter, PM10 and PM2.5, concentrations from eight air quality monitoring stations in Peninsular Malaysia. The duration of the time series is from July 2017 until June 2019. The results show that the best air quality predictive model is a model which is built using UCNN since it shows the highest accuracy for both time series PM10 and PM2.5. It is also shown that the forecast will have better accuracy if shorter past values are used as the input for the model.