Hybrid TCN-BiGRU model for accurate air quality index prediction

Accurate Air Quality Index (AQI) forecasting is critical for pollution control and public health protection, yet traditional methods struggle to capture long-range temporal dependencies and bidirectional context in complex atmospheric dynamics. We propose TCN-BiGRU, a novel deep learning framework that integrates Temporal Convolutional Network (TCN) and Bidirectional Gated Recurrent Unit (BiGRU) through a cascaded architecture with residual connections. The TCN component extracts multi-scale temporal features via dilated convolutions with exponentially increasing receptive fields, while BiGRU processes these enriched features bidirectionally to capture both past and future contextual dependencies. We evaluated TCN-BiGRU against 13 baseline models and 3 state-of-the-art models from 2025 across 16 seasonal-regional scenarios using datasets from Beijing (2022 and 2023), Guangzhou (2023), and Shanghai (2023). TCN-BiGRU achieves RMSE improvements over the strongest 2025 SOTA model in all 16 seasonal-regional scenarios, with gains up to about 17% and R² values consistently exceeding 0.876, while being 3.2-5.5× faster in inference than these SOTA models. Ablation studies confirm 25.5% performance improvement through synergistic effects of the cascaded architecture. Cross-year validation shows 8.3% performance degradation, while cross-city transfer learning achieves 74.2-82.7% of full local training performance with only 20% target data. These results demonstrate superior accuracy-efficiency trade-off and robust generalization across diverse geographic and climatic conditions. The source code is available at https://github.com/ProfYangPaperCode/TCN-BiGRU-for-Predicting-AQI.