Abstract

Air pollution is an issue of great concern. PM2.5 is the most dangerous pollutant out of all the pollutants. A large number of missing values is present in multivariate pollution data. This makes the prediction of PM2.5 concentration very difficult. Traditional approaches to deal with missing values includes mean imputation, median imputation, case deletion, matrix factorization-based imputation, case deletion, etc. All these methods are not capable of modelling the temporal dependencies and complex distributed nature of multivariate pollution data. We have dealt with missing values by mean imputation, median imputation, Generative Adversarial Network imputation and Conditional Tabular Generative Adversarial Network. We have used Artificial Neural Network to predict PM2.5 concentration. The results reveal that for time series data, imputation by median model has performed waste. No significance improvement is seen in prediction of PM2.5 concentration using GAN and CTGAN model if we compare it with imputation by mean model. For time series data, more research work needs to be done with GAN and its variants so that better results can be achieved