Deep Learning for AQI Prediction Using Multiple Feature Vectors: A Case Study of Colaba and Deonar Station

Abstract

Air pollution has been a vital global health challenge and is one of the major contributors towards mortality. Air quality monitoring, understanding and predicting are important for effective public health strategies. The goal is to design and compare different deep learning models to forecast air quality levels based on different input sets of features. The dataset consists of pollutant and meteorological parameters obtained from two monitoring stations in Mumbai: Colaba and Deonar. The proposed models are Recurrent Neural Network, Long Short-Term Memory, Bidirectional Long Short-Term Memory, and a hybrid model that connects one-dimensional convolutional layers with Long Short-Term Memory. The models were trained on three sets of features: pollutant-only, meteorological-only, and combined. Model accuracy was determined by root mean square error, coefficient of determination, mean absolute percentage error, and explained variance score. The best result for Colaba was attained by the hybrid model with the lowest prediction error and highest accuracy using the combined feature set. Deonar saw the Long Short-Term Memory model working best on the meteorological feature set alone. In both stations, Bidirectional Long Short-Term Memory models performed consistently irrespective of feature sets. The addition of meteorological data enhances prediction accuracy greatly. The use of Colaba station data provided more accurate predictions than using Deonar indicates the value of data quality and need of location-based modelling strategies. This research verifies that deep learning models, particularly hybrid and bidirectional architectures, serve as useful methods for air quality forecasting.