Evaluation of a transformer-based model for the temporal forecast of coarse particulate matter (PMCO) concentrations

Accurate forecasting of coarse particulate matter (PMCO) concentrations is crucial for mitigating health risks and environmental impacts in urban areas. This study evaluates the performance of a transformer-based deep learning model for predicting PMCO levels using 2022 data from four monitoring stations (BJU, MER, TLA, UIZ) in Mexico City. The transformer model's forecasting accuracy is assessed for horizons of 12, 24, 48, and 72 hours ahead and compared against conventional autoregressive integrated moving average (ARIMA) and long short-term memory (LSTM) models. Error metrics including root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) are employed for evaluation. Results demonstrate the transformer model's superior performance, achieving the lowest error values across multiple stations and prediction horizons. However, challenges are identified for short-term forecasts and sites near industrial areas with high PMCO variability. The study highlights the transformer model's potential for accurate PMCO forecasting while underscoring the need for interdisciplinary approaches to address complex air pollution dynamics in urban environments.