A Comprehensive Analysis of the Triple-Hybrid Metamodel of MLP-PSO-ARIMA for Forecasting the FDSD Index: A Case Study of Khuzestan Province

This study aims to evaluate the performance of the triple-hybrid metamodel MLP-PSO-ARIMA in forecasting the frequency of dust storm days (FDSD) index across seven selected stations in Khuzestan Province during a 50-year statistical period (1970–2019). The results of the proposed triple-hybrid metamodel were compared against the standalone MLP and ARIMA models, as well as the hybrid models MLP-PSO, ARIMA-PSO, and MLP-ARIMA, using performance metrics including R, RMSE, MAE, and NS. All the tested models demonstrated their highest accuracy during the first and second seasonal combinations. Accordingly, it was concluded that utilizing data from one or two preceding seasons yields more accurate predictions of the FDSD index for subsequent seasons in Khuzestan Province, whereas incorporating data from the third and fourth seasons does not enhance forecasting performance. Moreover, the multilayer perceptron (MLP) neural network outperformed the Box-Jenkins ARIMA model in predicting dust storm events in the region. While combining the MLP and ARIMA models improved the accuracy compared to their standalone counterparts, the improvement was not statistically significant. In contrast, the proposed triple-hybrid metamodel exhibited a statistically significant enhancement in accuracy over the dual-hybrid models.