Overcoming the Persistence Barrier in Air Quality Forecasting: A Zero-Initialized Residual Bi-GRU Framework

Articles in Press

Document Type : Original Article

Authors

Kimia Shirini ¹

Sina Samadi Gharehveran ²

Erfan Mohammadi ¹

Nima Saeedi ²

¹ Faculty of Multimedia, Tabriz Islamic Art University, Tabriz, Iran

² Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran

10.22067/cke.2026.98906.1192

Abstract

Accurate forecasting of short-term PM_2.5 concentrations is crucial for public health management but remains challenging due to the stochastic nature of atmospheric pollutants. Conventional neural architectures often fall short of surpassing basic persistence benchmarks, primarily because they tend to learn simplistic identity transformations when dealing with data characterized by strong temporal dependency. In order to overcome these barriers, we develop a specialized framework termed Zero-Initialized Residual Bidirectional GRU (ZIR-BiGRU), designed to enhance predictive performance. Unlike conventional approaches that model absolute pollution levels directly, the proposed architecture focuses on learning the non-linear residuals—the deviations from the current state—while utilizing a specialized zero-initialization strategy to ensure training stability starting from a persistence-equivalent state. The model was evaluated on real-world hourly air quality data, demonstrating robust performance across a 24-hour forecast horizon. Experimental results reveal that the ZIR-BiGRU significantly outperforms the persistence baseline, achieving a 20.60% improvement in the Coefficient of Determination (R^2) and a 4.37% reduction in Root Mean Squared Error (RMSE). Furthermore, the Diebold-Mariano test confirms the statistical significance of these improvements (p<0.05). Horizon analysis indicates a critical crossover point at h=6, beyond which the proposed model consistently dominates the baseline, proving its superior capability in capturing long-term dependencies and sudden fluctuations in air quality.

Keywords

Air Quality Forecasting

Prediction

Deep Residual Learning

Zero-Initialization

Bidirectional GRU

Time Series Analysis

Subjects

• Special Issues

---