Comparing the Performance of SARIMA and Holt-Winters Time Series Models With Artificial Intelligence Methods in Forecasting Dust Storms (Case Study: Sistan and Baluchestan Province)

The impact of the dust phenomenon in Iran is so great that it has involved more than half of the country's provinces in some way with the problems and limitations of this natural phenomenon, which in addition to environmental impacts, has disrupted the implementation of sustainable national development plans and so far it has had and will have many negative consequences. The increase in dust storms in recent years in the east and southeast of the country, especially in Sistan and Baluchestan province, and consequently the decrease in air quality in these areas, has doubled the importance of Forecasting this phenomenon. On the other hand, most domestic studies in this field are related to the process of small-scale dust phenomena, synoptic studies, and its satellites. Therefore, considering that this phenomenon has had adverse effects and negative consequences in the social, economic, and health fields of the people, it is necessary to study, forecast, and measure its relationship with climate variations.

This study aimed to compare the performance of SARIMA and Holt-Winters time series models with artificial intelligence methods including neural networks based on radial base functions (RBF) and adaptive neural-fuzzy inference system (ANFIS) to forecast the frequency of dust storm days (FDSD) in the next season. For this purpose, hourly dust data and codes of the World Meteorological Organization were used in five synoptic stations in Sistan and Baluchestan province with a statistical period of 25 years (1990-2014). The observations of meteorological phenomena are recorded once every three hours, a total of eight times a day. In these observations, the visual phenomena of climate are defined according to the guidelines of the World Meteorological Organization in 100 codes (00-99), in which 11 codes are used to record and report the phenomenon of dust in different meteorological stations. Following the time series of days with dust storms, the FDSD index was forecasted using four methods SARIMA, Holt-Winters, RBF, and ANFIS.

According to the results of the time series, the FDSD index in Saravan, Khash, Iranshahr, and Zahedan stations has relatively small variations that are scattered throughout the time series, but with the increase in the number of dust days in Zabol station, the scattering of the variations has decreased and its intensity has increased. Also, the peak values of dust are concentrated next to each other, which indicates the occurrence of successive dust storms in this station from 2000 onwards. As can be seen in the ACF and PACF diagrams of the studied stations, significant time intervals indicate the correlation between the time values that make it possible to modeling and forecasting future values (next season) of the FDSD index for all five stations studied. According to the functions of partial autocorrelation and autocorrelation, the range of change of attraction and the moving average was determined, and using the appropriate evaluation criteria, the best time series model was extracted for each station. In the Dickey-Fuller test, the significance level was considered to be P-Value < 0.05. According to the test, only the time series of Zabol stations is unstable, which confirms the results of the ACF and PACF diagrams of the studied stations. The results showed that the ANFIS method performed better than other methods in all study stations; So that in this method, the evaluation criteria of R, RMSE, MAE, and NS are varied from 0.72, 0.57, 0.42 and 0.71 to 0.95, 0.51, 0.40 and 0.96, respectively. Also, the average frequency of days with dust storm on a seasonal scale in the studied stations varied from 1.06 to 7.11, respectively, so that with increasing FDSD index in the stations, the forecasting accuracy of all methods increased. In the SARIMA time series model, the correlation coefficient (R) between the observed and forecasted values of the FDSD index was increased from 0.64 to 0.79. For Holt-Winters, RBF, and ANFIS methods, the R-value also varied from 0.70 to 0.87, 0.69 to 0.92, and 0.72 to 0.95, respectively. Also, based on the results of the observed and forecasted values, with the increase of FDSD index in the studied stations (progress from Saravan station to Zabol station), the relationship between observed and forecasted values in all methods (time series models and artificial intelligence methods) find more compatibility with the Semi-constructor of the first quarter. The results of the Z test also showed that the assumption of zero-based on the mean equal of the time series of observed and forecasted values of the FDSD index, in none of the studied stations based on ANFIS and RBF methods at 1% error level and based on SARIMA and Holt-Winters time series models are not rejected at the 5% error level.

The results showed that with the decrease in the frequency of days with dust storms in Saravan and Khash stations, the Holt-Winters time series model showed almost the same and higher performance than the RBF method, which indicated the high capability of this model to forecast low values FDSD index. The results also showed that the SARIMA time series model compared to other forecasting methods did not have a high ability in forecasting the FDSD index in any of the studied stations. Also, despite the low frequency of days with dust storms in Iranshahr station compared to Zahedan station, all FDSD index forecasting methods have better performance and more accurately than Zahedan station based on evaluation criteria, which can be searched due to the presence of a complete series without FDSD index termination at Iranshahr station. The results of this study can be useful in forecasting and managing the consequences of dust storms in the study areas. On the other hand, in forecasting the FDSD index in Sistan and Baluchestan province, the optimal predictor model has been complex. For all of the stations studied, the model that used three or four steps of the predictive delay was recognized as the best predictor model. Therefore, particles leftover from previous storms could be an important reason for the impact of the last few seasons’ storms on the formation of dust storms in future seasons.