Development and application of a hybrid long-short term memory - three dimensional variational technique for the improvement of PM2.5 forecasting.

Abstract The current state-of-the-art three-dimensional (3D) numerical model for air quality forecasting is restricted by the uncertainty from the emission inventory, physical/chemical parameterization, and meteorological prediction. Forecasting performance can be improved by using the 3D-variational (3D-VAR) technique for assimilating the observation data, which corrects the initial concentration field. However, errors from the prognostic model cause the correction effects at the first hour to be erased, and the bias of the forecast increases relatively fast as the simulation progresses. As an emerging alternative technique, long short-term memory (LSTM) shows promising performance in air quality forecasting for individual stations and outperforms the traditional persistent statistical models. In this study, a new method was developed to combine a 3D numerical model with 3D-VAR and LSTM techniques. This method integrates the advantage of LSTM, namely its high-accuracy forecasting for a single station and that of the 3D-VAR technique, namely its ability to extend improvement to the whole simulation domain. This hybrid method can effectively improve PM2.5 forecasting for the next 24 hours, relative to forecasting with the 3D-VAR technique which uses the initial hour concentration correction. Results showed that the root-mean-square error and normalized mean error were decreased by 29.3% and 33.3% in the validation stations, respectively. The LSTM–3D-VAR method developed in this study can be further applied in other regions to improve the forecasting of PM2.5 and other ambient pollutants.