Spatio-temporal long short-term memory neural network with seasonal-trend decomposition for ambient air pollutant forecasting

Ambient air pollution poses a significant threat to human health and quality of life. Accurate forecasting is crucial for mitigating these risks and guiding preventive actions. However, precise prediction remains challenging due to the non-stationary nature and complex spatio-temporal relationships of outdoor air pollutant dynamics. In this study, a novel model called the Seasonal-Trend Spatio-Temporal Long Short-Term Memory (ST²-LSTM) neural network is proposed for prediction. First, the seasonal-trend decomposition based on locally weighted scatterplot smoothing (STL) method is used to decompose air pollutant into trend, seasonal, and remainder components for each monitoring station. Then, the corresponding components from each station are combined to form spatio-temporal trend, seasonal, and remainder components to better learn spatial relationships in the next step. For the spatio-temporal components, the novel Spatio-temporal LSTM with attention model (AttenSTLSTM) is adopted. Notably, considering the stochastic nature of the spatio-temporal remainder component, a technique called reversible instance normalization (RevIN) is employed to correct distribution shifts. Finally, the predicted values of the spatio-temporal components are summed to obtain the final predicted values. Experimental evaluations are conducted in the real-world Jing-Jin-Ji area. Compared to eight baseline models, the ST²-LSTM model demonstrates superior performance across various stations and metrics. An ablation experiment is also conducted with six variants to verify the usefulness of each module in the proposed model. These findings highlight that the ST²-LSTM model offers higher prediction accuracy and stronger spatio-temporal modeling capabilities for accurate ambient air pollutant forecasting.