مقایسة عملکرد مدل‏ های سری زمانی sarima و holt-winters با روش‏ه ای هوش مصنوعی در پیش‏ بینی طوفان‏ های گردوغبار (مطالعة موردی: استان سیستان و بلوچستان)

هدف از این پژوهش مقایسة عملکرد مدل‏های سری زمانی sarima و holtwinters با روش‏های هوش مصنوعی شامل شبکة عصبی مبتنی بر توابع پایة شعاعی (rbf) و سیستم استنباط عصبی فازی تطبیقی (anfis) به‏منظور پیش‏بینی فراوانی روزهای همراه با طوفان گردوغبار (fdsd) در فصل آتی است. بدین منظور، از داده‏های ساعتی گردوغبار و کدهای سازمان جهانی هواشناسی در پنج ایستگاه سینوپتیک استان سیستان و بلوچستان با طول دورة آماری 25ساله (19902014) استفاده شد. نتایج نشان داد روش anfis، نسبت به سایر روش‏ها، بهترین عملکرد را داشت و معیارهای ارزیابی r، rmse، mae، و ns آن به‌ترتیب از 72/0، 57/0، 42/0، و 71/0 تا 95/0، 51/0، 40/0، و 96/0 متغیر بود. همچنین، با افزایش شاخص متوسط fdsd در ایستگاه‏ها (از 06/1 تا 11/7)، دقت پیش‏بینی همة روش‏ها افزایش داشت. بر همین اساس، در سری زمانی sarima، ضریب همبستگی بین مقادیر مشاهداتی و پیش‏بینی‏شده شاخص fdsd از 64/0 به 79/0 افزایش یافت. برای روش‏های holtwinters، rbf، و anfis مقدار نیز ضریب همبستگی به‌ترتیب از 70/0 تا 87/0، 69/0 تا 92/0، و 72/0 تا 95/0 متغیر بود. درمجموع، با مقایسة روش‏های مورد استفاده، روش هوش مصنوعی anfis بهترین و مدل‏های سری زمانی sarima و holtwinters بدترین عملکرد را داشتند.

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

IntroductionThe impact of dust phenomenon in Iran is so great that it has involved more than half of the country’s provinces with problems and limitations one way or another. In addition to environmental impacts, it has disrupted implementing sustainable national development plans, having already brought about many negative consequences. Increased number of dust storms in recent years in the east and southeast of the country, especially in Sistan and Baluchestan Province, and consequently the decrease of air quality in these areas, has doubled the importance of forecastingthis phenomenon. On the other hand, most domestic studies in this field are related to the process of smallscale dust phenomena, synoptic studies, and its satellites. Therefore, considering that this phenomenon has had adverse effects and negative consequences on social, economic, and healthcare aspects of human life, it is necessary to study, forecast, and measure its correlation with climate variations.Materials and MethodsThis study aimed at comparing the performance of SARIMA and HoltWinters time series models with artificial intelligence methods, including neural networks based on Radial Base Functions (RBF) and Adaptive NeuralFuzzy Inference System (ANFIS) to forecast the Frequency of Dust Storm Days (FDSD) in the next season. For this purpose, it used World Meteorological Organization codes as well as hourly dust data from five synoptic stations in Sistan and Baluchestan province with a statistical period of 25 years (19902014). The observations of meteorological phenomena were recorded once every three hours, making it eight times a day in total. In these observations, the visual phenomena of the climate were defined according to the guidelines of the World Meteorological Organization in 100 codes (0099), in which 11 codes were 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, HoltWinters, RBF, and ANFIS.Results and DiscussionAccording to the time series’ results, the FDSD index in Saravan, Khash, Iranshahr, and Zahedan stations displayed relatively small variations that were scattered throughout the time series, but with more dusty days at Zabol station, the variations’ scattering decreased and its intensity increased. Also, the peak values of dust were concentrated in close vicinity of one another, indicating the occurrence of successive dust storms at this station from 2000 onwards. As can be seen in ACF and PACF diagrams of the studied stations, significant time intervals indicate a correlation between the time values that make it possible to model and forecast the future values (next season) of the FDSD index for all five stations studied. The range of attraction change and moving average got determined according to partial autocorrelation and autocorrelation functions and using appropriate evaluation criteria, the best time series model was extracted for each station. In the DickeyFuller test, the significance level was considered to be PValue < 0.05. According to this test, only the time series of Zabol Station was unstable, confirming the results of ACF and PACF diagrams of the studied stations. Results showed that ANFIS Method performed better than other methods in all studied stations. Thus, in this method the evaluation criteria ranged from 0.72 to 0.95 for R, from 0.57 to 0.51 for RMSE, from 0.42 to 0.40 for MAE, and from 0.71 to 0.96 for NS. Also, the average frequency of days with dust storms on a seasonal scale varied from 1.06 to 7.11, allowing forecasting accuracy of all methods to increase as the FDSD index mounted. In SARIMA time series model, the correlation coefficient (R) between the observed and forecasted values of the FDSD index rose from 0.64 to 0.79. As for HoltWinters, RBF, and ANFIS methods, this value varied between 0.70 and 0.87, 0.69 and 0.92, and 0.72 and 0.95, respectively. Moreover, based on the results of the observed and forecasted values, the greater the FDSD index in the studied stations (progress from Saravan to Zabol Station), the more the compatibility of observed and forecasted values in all methods (time series models and artificial intelligence methods) with Semiconstructor of the first quarter. The results from the Z test also proved the assumption that stated that zerobased on the mean equal of the time series of FDSD index’ observed and forecasted values were not rejected in none of the studied stations, according to ANFIS and RBF methods at 1% and SARIMA and HoltWinters time series models at 5% error level.ConclusionThe results showed that as the frequency of days with dust storms in Saravan and Khash stations got less, the HoltWinters time series model showed an almost similar and better performance than the RBF method. This indicates this model’s high capability to forecast low values of FDSD index. The results also showed that the SARIMA time series model, compared to other forecasting methods, did not show a high ability to forecast FDSD index at any of the studied stations. Furthermore, despite the low frequency of days with dust storms at Iranshahr station, compared to Zahedan station, all FDSD index forecasting methods had a better performance and higher accurately than Zahedan station, based on the 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 for forecasting and managing the consequences of dust storms in the studied areas. On the other hand, when 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 cause for the impact of the last few seasons’ dust storms on the formation of new ones in the coming seasons.