Air quality prediction : a case study of the Maltese Islands

Abstract

In recent years, air quality has become an increasingly important factor in the wellbeing of humans. In the last decade, the European Union has enforced member countries to set up air monitoring stations to record different pollutant concentrations. In this regard, various research based on predicting pollutant concentration using deep learning techniques has been published, due to their capability of recognising patterns in the data. The domain of air quality comes with its complexities, especially when the sensors that are used to collect pollutant concentration get miscalibrated or stop functioning. This research focuses on understanding the different factors that contribute to air pollution and then uses those factors to predict the Air Quality Index (AQI). Since Malta does not have its own AQI, we also discuss different indices that are used in European countries. The data which is collected from the air monitoring stations comes out raw and does not contain the calculated AQI. The importance of obtaining such index arises from the categorisation of different levels of pollutant concentration. People are much more likely to understand a single numerical index rather than the concentration of pollutants themselves. To predict the air quality, we use the OSEMN framework to collect, transform, predict and visualise the air quality data. Different imputation techniques to handle missing data are evaluated and K-Nearest Neighbour (KNN) is selected to impute the dataset, where applicable. To predict the air quality: the CNN, LSTM and CNN-LSTM models are used to perform multivariate time series prediction. Four different datasets collected from the air monitoring stations located in Zejtun, Attard, Gharb and Msida, are used as the training sets for the deep learning models. The Convolutional Neural Network - Long-Short Term Memory (CNN-LSTM) model achieves the best score out of all the models, but performs the best when using the Attard dataset, obtaining a Root Mean Squared Error (RMSE) of 9.7 and a Mean Absolute Error (MAE) of 5.9. We use the results to calculate the Air Quality Index for Health (AQIH) and then visualise this index on a map of the Maltese Islands. This visualisation represents the areas in the Maltese Islands with the least and the most pollution using a colour coded heatmap.