Daily PM۲.۵ prediction and pollution episode detection using MLP neural networks across urban monitoring sites with varied land uses

This research presents a data-driven approach for forecasting daily PM۲.۵ concentrations using a multilayer perceptron (MLP) neural network across three urban monitoring sites in Mashhad, Iran—Sajjad, Torogh, and Vila—each reflecting a unique land-use profile. The study utilized daily datasets collected from ۲۰۱۸ to ۲۰۲۳, and a dedicated MLP model was trained for each station. Various training algorithms were assessed to identify the most suitable configuration, with model complexity fine-tuned by adjusting the number of neurons in the hidden layer. Key input features included meteorological variables from the preceding day (such as wind speed, ambient temperature, precipitation, solar radiation, and relative humidity), the previous day's PM۲.۵ concentration, and calendar-based temporal factors. To improve the network’s predictive capability and prevent overfitting, data normalization and early stopping strategies were applied. The best predictive performance was recorded at the Sajjad station, where the model achieved an R۲ value of ۰.۷۹ and an MAE of ۶.۷۷ µg/m۳. While the Torogh station yielded moderate predictive accuracy, the Vila station exhibited weaker performance. The models demonstrated strong capability in identifying pollution episodes, with true positive rates between ۶۶% and ۷۴%, and a minimum false alarm rate of ۰.۱۸ at the Sajjad station. Spatial disparities in model performance were attributed to localized environmental and climatic factors, including terrain variation and emission source intensity. Overall, the findings confirm the potential of MLP-based models as practical tools for daily air quality prediction and support their integration into urban pollution alert systems. .

کلیدواژه ها:

PM۲.۵ concentration forecasting ، Artificial neural networks ، Air pollution ، urban monitoring station ، meteorological data