Forecasting daily PM2.5 concentrations in Wuhan with a spatial-autocorrelation-based long short-term memory model

Abstract

Accurate daily air pollution forecasts play a pivotal role in enabling government to implement timely emergency responses and helping alert individuals sensitive to air pollution to take preventive measures. The atmospheric continuity fosters spatial correlations among air pollutants at various locations, which is a factor frequently overlooked in contemporary research focused on harnessing data-driven models for air quality prediction. Therefore, this study proposed a Spatial-Autocorrelation-based Long Short-Term Memory (SALSTM) model for the daily forecasting in Wuhan, Hubei Province, China. Using a multivariate prediction approach with daily air pollution data and meteorological data from Wuhan, as well as air pollution data from surrounding cities, from 2021 to 2022 as input, the model was applied for projecting the daily PM2.5 for Wuhan during the year 2023 and conducting accuracy cross-validation. The results were compared with a univariate prediction approach utilizing the Autoregressive Integrated Moving Average (ARIMA) model and the original Long Short-Term Memory (LSTM) model. Furthermore, this study utilized Dynamic Time Warping (DTW) for feature selection in multivariate prediction, comparing the accuracy of prediction results before and after feature selection. Experimental results indicated that the SALSTM model, incorporating the DTW method, achieved a Root Mean Squared Error (RMSE) of 6.92 μg/m3, a Mean Absolute Error (MAE) of 4.07 μg/m3 and a coefficient of determination (R2) of 0.95. Compared to the univariate forecasting method, the three accuracy metrics RMSE, MAE, and R2 have improved by 54.74%, 58.68%, and 37.68%, respectively. In comparison with the original LSTM, the improvement is 23.79%, 30.90%, and 4.40%. In conclusion, the SALSTM model established in this study demonstrates accurate daily forecasting of PM2.5 concentrations.