Evaluating the Efficiency of Hybrid Metamodels of Machine Learning and Box Jenkins in Order to Model Dust Storms (Case Study: Khuzestan Province)

The impact of dust phenomenon in Iran is so vast that it has involved more than half of the country's provinces in some way with the issues and limitations of this natural phenomenon. In addition to the environmental effects, it has disrupted the implementation of national sustainable development plans and so far, it has had many negative consequences. This research tries to present a new hybrid model using artificial intelligence hybrid metamodels as well as Box Jenkins hybrid metamodels to predict and model the FDSD index (frequency of days with dust storms), in seven synoptic stations of Khuzestan province with length The statistical period has been 40 years (1981-2020). The hybrid prediction algorithms used in this research include W-ANFIS, AF-SVM, ARIMA-NARX and SARIMA-SETAR. The prediction results showed that the decrease in the performance of hybrid models to predict the FDSD index has a direct relationship with the decrease in the frequency of days with dust storms. So that the correlation coefficient for experimental data in AF-SVM and W-ANFIS hypermodels from 0.991 and 0.985 to 0.985 and 0.958, respectively, and Nash Sutcliffe coefficient has also decreased from 0.977 and 0.960 to 0.973 and 0.952, respectively. Also, the RMSE coefficient from Abadan station to Dezful for the two metamodels from 0.135 and 0.151 to 0.140 and 0.179 respectively, And the MAE coefficient has also increased from 0.054 and 0.068 to 0.060 and 0.093, respectively. Correlation coefficient for test data in Box Jenkins SARIMA-SETAR and ARIMA-NARX hypermodels also from 0.967 and 0.951 to 0.958 and 0.941 respectively and the Nash Sutcliffe coefficient has also decreased from 0.945 and 0.923 to 0.938 and 0.913, respectively, which indicates the weakening of the performance of hybrid metamodels with the decrease in the frequency of dust storms in Khuzestan province. Also, by fitting four hybrid hypermodels on the FDSD index, it was shown that AF-SVM hybrid hypermodel had better performance than other methods. In a way, in all studied stations, the correlation coefficient and Nash-Sutcliffe coefficient are higher and the root mean square error coefficient and the mean absolute value of the error are lower, which shows the superiority of this hybrid meta-model over other meta-models for predicting the FDSD index in Khuzestan province. The results of this study can be used to model dust storms in other western regions of the country.