PM2.5 Forecasting in an Urban Environment A Multi-Horizon Comparison of Statistical and Deep Learning Models in Santiago, Chile

Abstract

Forecasting fine particulate matter (PM2.5) concentrations is an important task in urban environmental management because of its effects on human health and its association with critical air pollution episodes. This study presents a controlled univariate multi-horizon comparison between classical statistical models and deep learning for daily PM2.5 forecasting at an urban monitoring station in Santiago, Chile. The analysis used an official time series from the Sistema de Información Nacional de Calidad del Aire (SINCA) covering 2010–2025. Four approaches were evaluated: a weekly Seasonal Naïve baseline, ARIMA, SARIMA, and a Long Short-Term Memory (LSTM) network. Forecasts were assessed at 1, 7, 14, 21, and 28 days ahead using a chronological train-validation-test split and temporally consistent model selection and evaluation procedures. Performance was measured with RMSE and R2. The results showed that LSTM achieved the best overall performance across all forecasting horizons, with RMSE values ranging from 7.61 to 12.98, while the Seasonal Naïve baseline showed the weakest results. LSTM also obtained the highest explanatory performance, consistently outperforming ARIMA, SARIMA, and the baseline. These findings indicate that even under a strictly univariate formulation, model choice has a substantial effect on PM2.5 forecasting accuracy.