Urban PM2.5 Forecasting with Machine Learning: A Comparative Analysis

Published comparisons of PM2.5 prediction models reach different conclusions about the value of deep learning. This review examines eleven empirical studies and one methodological review published between 2018 and 2024 to assess those differences. The papers are compared by prediction target, input data, and the alternatives included in each experiment. Support vector regression and linear regression performed well in a Nottingham study using features extracted from recent concentration trends. Other studies reported lower errors with combined convolutional and recurrent networks or Transformers. In these systems, however, changes to the predictor often accompanied changes to feature selection or the representation of spatial information. The reported improvement therefore cannot always be attributed to network architecture alone. Forecast duration presents a further difficulty: estimating the next day's mean concentration and predicting an hourly peak require different evaluations. Conventional regression provides a useful baseline for these comparisons. A proposed replacement needs testing at the intended lead time, including pollution episodes, with inputs restricted to information available when the prediction is issued.