Air quality monitoring and mitigation through time series forecasting and stochastic optimization.

Abstract

Poor air quality poses significant threats to public health and environmental sustainability. To mitigate such risks, accurate air quality prediction is essential to inform intervention policies that effectively reduce pollutant levels. While past research has focused on forecasting air quality trends, this paper proposes a novel predict-then-optimize framework that integrates machine learning models with a two-stage stochastic programming model. Our approach first forecasts fine particulate matter (PM2.5) levels then leverages these predictions in an optimization model to identify mitigation strategies for cities in Ontario, Canada. In the prediction phase, we develop and evaluate multiple machine learning models, including Random Forest, XGBoost, LSTM, Stacked LSTM, and ensemble architectures. These models leverage meteorological, wildfire, and historical air quality data. The predictions from the best-performing model are then used as inputs to a two-stage stochastic programming model, which selects optimal intervention policies for different cities while considering uncertainty in pollutant levels and adhering to budget constraints. Extensive computational experiments demonstrate the ensemble model's superior predictive performance compared to all other forecasting models achieving an RMSE of 3.305. The results also highlight the effectiveness of the proposed stochastic programming model to identify mitigation policies that reduce PM2.5 levels in all cities, with the majority of cities falling below the recommended limit.