Advanced Hybrid Models for Air Pollution Forecasting: Combining SARIMA and BiLSTM Architectures

This study explores a hybrid forecasting framework for air pollutant concentrations (PM_10, PM_2.5, and NO₂) that integrates Seasonal Autoregressive Integrated Moving Average (SARIMA) models with Bidirectional Long Short-Term Memory (BiLSTM) networks. By leveraging SARIMA’s strength in linear and seasonal trend modeling and addressing nonlinear dependencies using BiLSTM, the framework incorporates Box-Cox transformations and Fourier terms to enhance variance stabilization and seasonal representation. Additionally, attention mechanisms are employed to prioritize temporal features, refining forecast accuracy. Using five years of daily pollutant data from Romania’s National Air Quality Monitoring Network, the models were rigorously evaluated across short-term (1-day), medium-term (7-day), and long-term (30-day) horizons. Metrics such as RMSE, MAE, and MAPE revealed the hybrid models’ superior performance in capturing complex pollutant dynamics, particularly for PM_2.5 and PM₁₀. The SARIMA combined with BiLSTM, Fourier, and Attention configuration demonstrated consistent improvements in predictive accuracy and interpretability, with attention mechanisms proving effective for extreme values and long-term dependencies. This study highlights the benefits of combining statistical preprocessing with advanced neural architectures, offering a robust and scalable solution for air quality forecasting. The findings provide valuable insights for environmental policymakers and urban planners, emphasizing the potential of hybrid models for improving air quality management and decision-making in dynamic urban environments.