A novel interpretable ozone forecasting approach based on deep learning with masked residual connections

Abstract

Air pollution is a growing threat, especially in low- and middle-income countries, causing over 4 million premature deaths annually. Ground-level ozone is a major concern, demanding accurate and interpretable prediction systems for effective public health management. However, existing time-series forecasting methods struggle to capture both linear and nonlinear dependencies in atmospheric data. This study introduces ResSelNet, a novel Residual Selection Network that integrates masked residual connections and embedded feature selection within a unified deep learning architecture. The model dynamically determines the optimal processing depth for each feature, allowing linear relationships to bypass nonlinear transformations while capturing complex patterns when necessary. Applied to five monitoring stations across Andalusia (Spain), ResSelNet consistently outperformed state-of-the-art baselines, achieving 8%–12% lower RMSE and MAE than LSTM and Transformer models. Beyond accuracy, the framework improves interpretability and robustness, revealing the hierarchical relevance of meteorological and pollutant variables. ResSelNet therefore offers an effective and explainable solution for multi-horizon environmental time-series forecasting.