Air pollution prediction with advanced preprocessing and deep ensemble learning

Abstract

The research presented investigates the impact of model selection and preprocessing techniques on air pollution prediction performance, particularly pertaining to achieving Sustainable Development Goals (SDGs). Accurate training of predictive models necessitates effective handling of missing or null values in environmental datasets. To address this challenge, Probabilistic Principal Component Analysis (PPCA) and the Extra Tree Regressor for data imputation are employed, followed by scaling using Robust Scaler, Min-Max Scaler, and Standard Scaler. A thorough comparison of these preprocessing methods revealed that PPCA is the most suitable choice for imputing missing values in air quality datasets, while the Robust Scaler provided the most reliable and accurate scaling. Additionally, Stochastic Gradient Descent (SGD) is integrated as an optimization technique to enhance model performance. The Weighted Average ensemble method, combining PPCA imputation and Robust Scaler, demonstrated superior predictive capabilities. This research highlights the potential for further improvements through additional ensemble techniques and model optimization strategies, opening avenues for future research focused on improving prediction precision and advancing the achievement of Sustainable Development Goals linked to environmental sustainability.