Authors
Angelo Casolaro, Vincenzo Capone, Francesco Camastra
Journal
Abstract
Nitrogen dioxide pollution is an ongoing and growing environmental issue that affects human health in developed Western countries. This study introduced a Bayesian attention-based deep neural network model for predicting ground-level nitrogen dioxide concentrations. The proposed model integrates the principles of the Bayesian neural network and the attention mechanism, enabling it to produce predicted values and their associated uncertainties, expressed as standard deviations. The proposed model was validated using 2020 data collected from 520 European Environmental Agency stations, located in Italy. The performance of the model was assessed using the mean absolute error.
Description
The article titled “Predicting Ground-Level Nitrogen Dioxide Concentrations Using the Bayesian Attention-Based Deep Neural Network” presents a novel deep learning approach for modeling air pollution dynamics.
The study introduces a Bayesian attention-based neural network that simultaneously predicts nitrogen dioxide (NO₂) concentrations and quantifies the uncertainty of its forecasts. By combining Bayesian inference with an attention mechanism, the model enhances both interpretability and reliability.
Validated on 2020 data from 520 monitoring stations across Italy, the approach achieved strong predictive performance, offering a valuable tool for air quality management and environmental policy support.