Evaluating deep learning time series models for PM2.5 forecasting across diverse horizons

Air pollution, particularly PM2.5, poses a major health challenge in urban areas such as Chengdu, China, where basin topography and intense emission sources exacerbate pollutant concentrations. This study evaluates four deep-learning time series algorithms-LSTM, CNN-LSTM, Transformer, and Transformer-LSTM-for PM2.5 forecasting, comparing univariate and four multivariate configurations incorporating auxiliary pollutants (CO, NO₂, O₃, SO₂) and meteorological factors (temperature, pressure, precipitation, wind speed). Using two years of daily data (November 2022-October 2024), models are assessed across monthly, seasonal, half-year, and annual horizons with complete and incomplete seasonal datasets. Results demonstrate: Transformer-LSTM yields superior performance with higher R² and lower MAE% and RMSE%, especially when augmented by meteorological factors over pollutants; complete seasonal training improves performance, while gaps exceeding three months between training and prediction reduce reliability due to evolving PM2.5 dynamics. These findings underscore meteorological integration, data-driven modeling, seasonal completeness, and timely prediction for pollution control for policymakers in Chengdu.