Evaluating the Performance of GRU-LSTM Hybrid Model in Predicting the Dust Storms Events (Case Study: Khuzestan Province in Southwest of Iran)

Understanding the frequency of dust storms in each area and being mindful of temporal-spatial variation of this event can help to monitor and reduce the damages induced by dust events. Due to the increasing development of metamodels and their combination with optimization algorithms used to model and predict hydrological variables, machine learning models due to high accuracy in forecasting, in the form of a black box, have received a lot of attention. Therefore, in the present study, a hybrid approach is proposed to predict the Frequency of Dust Storm Days (FDSD) on a seasonal scale, which uses a combination of Lang Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) neural networks. In this study, the performance of the proposed hybrid model was compared with a neural network based on Radial Basis Functions (RBF) and Support Vector Machine (SVM). For this purpose, hourly dust data and codes of the World Meteorological Organization were used on a seasonal scale with a statistical period of 30 years (1990-2019) for seven synoptic stations in Khuzestan province. The results of the evaluation criteria in the training and testing Stages showed that the GRU-LSTM hybrid model offered better performance than other models used to predict the frequency of days with dust storms; The proposed hybrid model with correlation coefficient (R) of 0.905-0.988, Root Mean Square Error (RMSE) of 0.313-0.402 day, Mean Absolute Error (MAE) of 0.144-0.236 day, and Nash-Sutcliffe (NS) of 0.819-0.903 had better performance compared to the other models used in predicting the FDSD index. In general, comparing the models used in this study, the GRU-LSTM hybrid method and the SVM model, respectively, provided the best prediction skills. As a result, application of the proposed hybrid model can be used as a suitable tool to predict the FDSD index and adopting management decisions to reduce the dust storms damages in the study area.