Multivariate Hourly Air Quality Forecasting with MES-LSTM as the Core Framework

Accurate air quality forecasting plays an important role in supporting environmental management, protecting public health, and enabling timely early-warning interventions. However, many existing forecasting approaches rely on univariate modeling and may not adequately capture interactions among multiple pollutants. This study therefore explores multivariate hourly air quality forecasting using Multivariate Exponential Smoothing--Long Short-Term Memory (MES-LSTM) as the primary forecasting framework. Experiments were conducted using the Beijing Air Quality dataset to predict carbon monoxide (CO), nitric oxide (NO), nitrogen dioxide (NO₂), ozone (O₃), and sulfur dioxide (SO₂). A standardized evaluation protocol was adopted for all models, including identical preprocessing procedures, chronological data partitioning, a 24-hour lookback window, and a one-hour forecasting horizon. The proposed MES-LSTM model was evaluated against Long Short-Term Memory (LSTM), Temporal Fusion Transformer (TFT), and a hybrid MES-LSTM--TFT architecture. The results show that MES-LSTM achieved the lowest forecasting errors for several pollutants, particularly CO, NO₂, and O₃, while remaining competitive for NO and SO₂. In contrast, the hybrid MES-LSTM--TFT model did not improve forecasting performance and generally produced larger prediction errors than the corresponding single-stage models. Additional robustness analyses, including five-seed experiments, persistence baselines, and multivariate coupling ablations, indicate that the effectiveness of pollutant interaction modeling depends on the target pollutant being predicted. Overall, the findings suggest that combining exponential smoothing with residual learning provides an effective and robust solution for short-term multivariate air quality forecasting. The study further demonstrates that MES-LSTM provides a strong and competitive forecasting framework under a unified evaluation protocol, while increased architectural complexity does not necessarily lead to improved predictive performance. Additional robustness analyses, including five-seed experiments, persistence baselines, and coupling ablations, were conducted to assess stability and reproducibility. Computational cost was not formally benchmarked and remains an avenue for future investigation.