Predicting missing value from multi-site time series atmospheric ozone data

Abstract

Air quality of this days is a vital issue for human health. But not only for human but also for the earth. Increasing of harmful pollutants in air plays a major for global warming. Increase in daily life temperature causes the significant increase of sea level. Based on IPCC’s 2021 report, city of India, Kolkata will be in underwater by 2030 unless a drastic change happens in climate change. Breathing in polluted air for years can cause of deadly disease, cancer. Many researchers actively started working on the air pollutant prediction model. Many of them worked on neural network-based models to predict the concentration of air pollutants. In this study, we will predict ozone concentration of a particular air quality monitoring station. We worked with five machine learning (Elastic Net, Decision Tree, Random Forest, SVR, LightGBM) and one deep learning (ANN) model and implementing five different feature selection methodologies (Radius based, Triangulation, Mutual Information, K-L divergence, Cluster based). We have used MAE (Mean Absolute Error) and RMSE (Root Mean Square Error) to analyse our model’s performance. LightGBM has given best result compared to all air pollutant prediction models.