مدل‌سازی و پیش‌بینی رواناب ماهانه در دامنه زمان (مطالعه موردی: حوضه آبریز قره‌سو) مقاله

در پژوهش حاضر برای مدل‌سازی رواناب ماهانه، از داده‌های چهار ایستگاه هیدرومتری پل توسکاستان، نهارخوران، غازمحله و سیاه‌آب در حوضه آبریز قره‌سو در یک دوره آماری 36 ساله استفاده شد. سپس بررسی همگنی سری داده‌ها با استفاده از آزمون چاو انجام شد. پس از مرتب‌سازی‌ داده‌ها، برای مدل‌سازی مقادیر رواناب ماهانه از چهار روش‌ باکس و جنکینز (SARIMA)، شبکه عصبی مصنوعی (ANN)، شبکه عصبی مصنوعی ـ فازی (ANFIS) و الگوریتم ژنتیک (GA) در ایستگاه‌های هیدرومتری منتخب استفاده شد. پس از مدل‌سازی مقادیر رواناب‌ ماهانه با استفاده از چهار مدل فوق، به پیش‌بینی تغییرات رواناب ماهانه در ایستگاه‌های هیدرومتری منتخب برای دوازده ماه آینده پرداخته شد و این امر با کمک نرم‌افزارهای Minitab، R و SPSS صورت گرفت. با توجه به نوع پراکنش مقادیر رواناب و وجود داده صفر، برای تثبیت واریانس از تبدیلlog (1+Yt) در مدل استفاده شد. در مرحله بعد، اعتبارسنجی مقادیر پیش‌بینی شده توسط مدل‌ها با استفاده از شاخص‌های MAD، RMSE و MAPE ارزیابی شد. نتایج نشان داد که در اکثر ایستگاه‌های هیدرومتری منتخب، مدل شبکه عصبی مصنوعی بهترین عملکرد را در بین چهار مدل مورد استفاده داشت. بعد از شبکه عصبی مصنوعی، شبکه عصبی مصنوعی ـ فازی دارای مناسب‌ترین عملکرد بود. روش باکس و جنکینز نیز با وجود اینکه در تشخیص روند تغییرات به صورت مناسب عمل کرده بود، در بین چهار مدل مورد استفاده عملکرد ضعیف‌تری را در پیش‌بینی مقادیر رواناب داشت.

1- Introduction Nowadays, demand for water is increasing especially in arid and semi-arid regions (e.g., Iran) due to population growth, economic development, higher standard of living, and changes in consumption patterns. Hence, optimal management of water resources in these areas is essential. Furthermore, climate change and increasingly extreme weather events have caused a surge in natural disasters (e.g., floods) over the past 50 years in arid and semi-arid regions. Thus, forecasting and modeling of runoff data is extremely necessary for planning and managing of water resources. Water flow forecasting plays a key role in flood reduction, reservoir optimization, and reservoir management. These models are mostly developed and applied for simulation and prediction. Therefore, different types of forecasting methods have been proposed over the decades including: Box and Jenkins (SARIMA), Artificial Neural Network (ANN), Adaptive Neuro-Fuzzy Inference Systems (ANFIS), and Genetic Algorithm (GA) models. Forecasting hydrological reactions invariably involves uncertainty. So far, numerous studies have been performed to improve the reliability and accuracy of hydrological forecasts, resulting in reduced risk error. Therefore, the main objective of current research was to use artificial intelligence methods consisting of ANN, ANFIS, GA, and SARIMA models to predict monthly runoff data and also select the best model for the efficient management of water resources in the Gharasou River basin. 2- Methodology Gharasou river basin with an area of 1624 square kilometers is located in the west of Golestan province and has an important role in providing water resources required in this province. In this research, to model and forecast the runoff process, the monthly runoff time series of 4 hydrometric stations of Pol-Tuskestan, Naharkhoran, Ghazmahale, and Siah-ab over Gharasou River basin were used for a period of 36 years (1982-2018). The time series homogeneity was examined using the Chow`s method. Runoff data are time dependent, so initially these data were arranged in time series. After sorting the data, four models consisting of Box and Jenkins (SARIMA), Artificial Neural Network (ANN), Adaptive Neuro-Fuzzy Inference Systems (ANFIS), and Genetic Algorithm (GA) models were used to forecast monthly runoff. To increase the prediction accuracy of other methods, the far time series of monthly runoff were first ignored based on to neural network method and then the number of effective years for modeling was determined. Later, the monthly runoff was forecasted for the next 12 months using four models consisting of SARIMA, ANN, ANFIS, and GA. Lastly, based on the forecasted values and using MAD, RMSE and MAPE indices, the accuracy and precision of SARIMA, ANN, ANFIS and GA models were compared. Modeling and forecasting were done using Minitab, R and SPSS software packages. 3- Results Based on the type of distribution of monthly runoff and the presence of zero data, log(1+Yt) conversion was used in the models to stabilize the variance. The results according to the autocorrelation diagrams revealed that the time series in all stations have seasonal trend with a period of 12 months. Then, the monthly runoff of the next 12 months was forecasted using four models including SARIMA, ANN, ANFIS and GA. Model validation results using three indicators of MAD, RMSE and MAPE revealed that the ANN model in the three hydrometric stations of Naharkhoran, Pol-Tuskestan and Siah-ab had the best performance. In these three hydrometric stations, after ANN model, the ANFIS model has been selected as the most suitable model. However, the performance of these two models has been very similar. In the Ghazmahaleh hydrometric station, two models of ANFIS and ANN had the best performance, respectively. In this study, it was also found that in four selected hydrometric stations, the GA model had a good performance after the two models of ANN and ANFIS. Although SARIMA model performed very well in identifying the trend of monthly runoff changes, it had the weakest performance among the methods. The forecast data using SARIMA model were overestimated compared to the actual data for March and April, but in other months, the forecast data using this model were relatively appropriate. 4- Discussion & Conclusions In this research, to model and predict the monthly runoff process, four models including SARIMA, ANN, ANFIS, and GA models were used for four selected hydrometric stations in Gharasou River basin. The results of model validation using three indicators of MAD, RMSE and MAPE showed that the ANN and ANFIS models had the best performance among the four models used. It was also found that time series of runoff data in hydrometric stations have undergone structural changes due to the physical and climatic alterations in the upstream of rivers. Therefore, the distant past of time series may cause deviations in modeling and forecasting results. To overcome this problem, making use of an algorithm to select the number of effective years in modeling and forecasting can be useful. Artificial neural network provides a suitable criterion for selecting the number of effective years due to its high accuracy in modeling and forecasting. Based on the effective years identified by this model, other models can be modified and provide more appropriate input data for forecasting from other models.