A Hybrid GRU-SVR Ensemble Framework for Hourly PM2.5 Forecasting and Adaptive Air Quality Management in Pakistan

Abstract

Precise short-term forecasting of PM2.5 levels is crucial for building effective early warning systems and reducing public health hazards, especially in countries like Pakistan, where the combination of temperature inversions and burning of agricultural waste leads to severe smog events during winter. Motivated by these challenges, proposed holistic PM2.5 forecasting framework based on a dataset comprising 2,193 daily air quality monitoring records. An empirical study is conducted on eight ML regression models, five DL sequence models (LSTM, GRU, BiLSTM, CNN1D, and windowed MLP), and two proposed ensemble models called AeroSynNet and AeroStackNet. The combination of GRU and SVR in AeroSynNet model is achieved via weighted averaging, whereas AeroStackNet combines GRU, CNN1D, SVR, and LightGBM, with Ridge Regression serving as the meta-modeller. From experimental results, it is evident that Linear Regression achieved best overall performance, closely followed by proposed AeroStackNet ensemble, with Linear Regression's RMSE equal to 3.866, MSE equal to 14.944, MAE to 2.903, and R2 value 1 of 0.9050 (AeroStackNet: RMSE 3.885, R2 0.9040). After correcting ensemble weight-tuning procedure to use held-out validation split rather than test set, proposed AeroSynNet framework no longer ranked first; the Wilcoxon Signed-Rank Test confirmed that Linear Regression, GRU, and AeroStackNet were statistically indistinguishable from one another, while AeroSynNet performed significantly worse than each of them. However, the added complexity of AeroStackNet provides a notable performance gain over the simpler AeroSynNet blend. Linear Regression itself matched or exceeded the accuracy of every Deep Learning and ensemble model at a fraction of the computational cost. From seasonal error analysis, it was observed that largest forecast errors occurred in winter, owing to smog. Collectively, study shows that a well-engineered Linear Regression baseline, together with proposed ensemble, provides an accurate, efficient, and statistically robust solution, whereas simpler two-model approach proved more sensitive to tuning of its weights.