Uncertainty-Aware Neural Forecasting of Air Quality in Semi-Arid Mexican Metropolises: A Case Study of the Guadalajara Metropolitan Area

Abstract

Air quality degradation in rapidly urbanizing regions involves complex, non-linear dynamics that frequently elude traditional deterministic models. To tackle this limitation, this research introduces a probabilistic forecasting and transfer station methodology. By reliably extracting patterns from official sensor data, our framework predicts the Air Quality Index (AQI) within the Guadalajara Metropolitan Area, Mexico. The approach utilizes NeuralProphet, a hybrid architecture that integrates deep Auto-Regressive Networks with seasonal decomposition. Based on hourly concentrations of six criteria pollutants (PM10, PM2.5, O3, NO2, SO2, CO) collected from four official monitoring stations between 2022 and 2024, a preprocessing pipeline aligned with US EPA standards was implemented. Optimization identified a 12-h lag window to forecast a 1-h future horizon. To evaluate the model’s generalization, the cross-station framework was trained on an aggregated signal from peripheral stations and tested on the industrial core of Miravalle. The framework was benchmarked against standard deep learning architectures (LSTM, GRU, RNN). While recurrent models exhibited high deterministic accuracy, NeuralProphet maintained robust predictive performance (R2≈0.995 on the held-out industrial test set) while additionally providing explicit seasonal interpretability and 80% prediction intervals via quantile regression. This quantification of uncertainty helps distinguish stable diurnal trends from sporadic hazardous events. By providing probabilistic forecasting ranges rather than deterministic point estimates, this transparent, data-driven tool offers a practical option for proactive public health interventions alongside black-box alternatives.