Iranian Water Research Journal

Abstract Evaporation is a physical phenomenon by which water particles enter the atmosphere in the form of vapor from the surface of water or the surface of wet soil through receiving solar energy. Evaporation process is an essential part of the hydrological cycle and it takes place continuously in nature. The importance of evaporation estimation in water resources and agriculture studies is undeniable. The evaporation pan is used as an indicator to determine the evaporation from the water level of lakes and reservoirs all over the world due to the ease of interpreting its data. Evaporation prediction is one of the important variables in planning and managing water resources. Various models are used to predict evaporation, such as stochastic time series models, experimental models, data mining models. Due to the fact that evaporation is affected by many parameters that are non-linear, non-stationary and random and occurs in a dynamic and complex manner, the use of experimental methods to predict evaporation that is based only on weather information such as temperature, relative humidity, wind speed and solar radiation are used, the accuracy is not enough. So far, various methods have been presented to calculate and estimate evaporation, such as the water balance method and experimental methods. These methods have flaws such as the inability to generalize to other places and different climatic conditions. Therefore, researchers tried to find a way to reduce the error as much as possible. The use of machine learning models are among these methods.

Materials and Methods

In this research, two techniques of support vector machine and network with long short-term memory will be used to estimate the amount of evaporation from the pan in Kohgiluyeh and Boyer-Ahmad Province, and the results will be compared with the actual evaporation values recorded in evapotranspiration stations and some experimental methods of evaporation estimation. In order to model and estimate the amount of evaporation, daily data of parameters of pan evaporation, rainfall, wind speed, maximum and minimum temperature, cloudiness, soil temperature, maximum and minimum relative humidity, average dew point temperature, sea level pressure and actual vapor pressure has been prepared from the regional water organization for 6 stations including Yasouj, Imamzadeh Jafar, Dehdasht, Dogonbadan, Siskhet and Likk. The length of the statistical data period varies from 39 years in Dogonbadan to 14 years in Likk. In this research, two SVM and LSTM models, as well as three experimental methods of Mayer, US Civil Engineering Organization (USBR) and Ivanov were used to estimate evaporation from pans. To use the models to estimate evaporation, first the data were randomly divided into training and testing stage. 70% of the data were considered as training data and 30% of the data were considered as test data. Finally, in order to evaluate the models, the statistical indices of explanation coefficient (R2), root mean square error (RMSE) and mean absolute value of error (MAE), model efficiency index or Nash Sutcliffe coefficient (N.S) and residual mass index (CRM) were used.

Results and Discussion

The results of the SVM model show that the error values of daily pan evaporation estimation in the SVM model in the training stage vary between 1.6 and 1.2 mm per day. The efficiency coefficient also shows that the model is in a satisfactory condition in terms of estimation and forecasting ability in the Siskhet station, in a good condition in the Yasouj station, and in a very good condition in other stations. LSTM model has weaker results than SVM model. The error values in this model are higher than the SVM model and vary between 2.2 and 3 mm per day. Also, the results of the efficiency coefficient of the model (N.S) also show that this model has not provided acceptable results for estimating pan evaporation in the two stations of Sisakht and Yasouj. In part of the models test, the LSTM model also showed a weaker ability than the SVM model in most of the stations. The LSTM model has provided better results than the SVM model only in Imamzadeh Jafar station. The best result among all the stations and the two investigated models was obtained at the Dehdasht station with an error of 1.6 mm per day and a Nash coefficient of 0.87 for the SVM model. Comparison of the results of the models with 3 experimental methods of Mayer, Ivanov and USBR also showed that, except for Ivanov's method and at Doganbadan station, none of the experimental methods could simulate the daily evaporation values well.

Conclusion

In summary, the results of this research show that the SVM model has estimated the northern and central regions of the province with an error between 15 and 18 percent, while the LSTM model estimates the same regions with a greater error and about 22 to 26 percent. In the south of the province, the results is almost reversed and the SVM model has an error of about 19 to 21 percent and the LSTM model has an error of between 13 to 21 percent. In the east of the province, the amount of error in the estimation of pan evaporation in the LSTM model is also higher than in the SVM model. In total, the comparison of the results of two models and three experimental methods of estimating evaporation from pans in Kohgiluyeh and Boyer-Ahmad provinces showed that LSTM model is the best model for Imamzadeh Jafar station and SVM model is the best model for the other 5 investigated stations. None of the three experimental models, provided acceptable results for the province. The results of this research show that, despite the emergence of powerful models that are capable of detecting complex and non-linear behaviors of variables based on mathematical relationships that can well recognize and simulate the pattern of data variability, the use of experimental methods to estimate pans evaporation is not recommended.

Abstract Abstract

Introduction:

Evaporation from water levels is one of the most important processes in meteorology and hydrology. Optimum use of available water resources is necessary due to the high costs of water storage in the dams. Every year, it is evaporate millions of cubic meters of fresh water from the dam’s water. On the other hand, evaporation from the pan helps to estimate the amount of the plant water requirement in areas where there is no lysimeter information. A wide variety of methods can be used to measure or estimate water use by plants. The evaporation pan method is a simple, inexpensive, and readily understandable way to estimate irrigated crop water use. It also requires little attention and can provide farmers with reliable estimates of plant water use. Pans provide a measurement of the integrated effect of radiation, wind, temperature, and humidity on the evaporation from an open water surface. The pan has proved its practical value and has been used successfully to estimate water use by observing the evaporation loss from a water surface and applying the crop coefficients to relate pan evaporation to irrigation requirement. For this reason, the study of evaporation from the pan has always been of interest to researchers. Modeling this important parameter can be very effective and helpful in accurately estimating of missing data.

Methods:

In this research, by using of the Gene Expression Programming model (GEP) and some time series models, including Exponential smoothing model, Holt’s model, ARIMA, Brown’s linear trend and Winters’ model, the daily evaporation from pans in Khuzestan province was evaluated. For this purpose, daily pan evaporation for 20 years (2000 to 2020) of three synoptic stations of Ahvaz, Abadan and Dezful were used. These stations have the longest and most complete data among the stations of this province. For modeling the pan evaporation data, 80% of the data was used for the training stage and the remaining 20% was used for the model testing stage. In this research, seven patterns were defined for data modeling. These patterns include the use of data from 20 months ago (M20), 16 months ago (M16), 14 months ago (M14), 10 months ago (M10), 6 months ago (M6), 4 months ago (M4) and 2 months ago (M2). Three evaluation statistics ware also used for evaluate the results of the models. This statistics are include the Root Mean Squared Error (RMSE), Nash- Sutcliffe coefficient (N.S) and Coefficient of Determination (R2). The model with the lowest root mean squared error value and the highest Nash- Sutcliffe coefficient and coefficient of determination was selected as the best model.

Results:

The evaluation of the results based on the 3 evaluation indices, showed that the patterns M14, M16 and M20 have the highest accuracy in modeling in both training and test stages. In other words, by increasing the memory of the model, the results show more accuracy and less error. Also, based on the evaluation statistics, the amount of simulation error in the cold months (October to March) was 28% lower than in the hot months (April to September). However, in all 12 months of the year, at least one suitable patterns with acceptable accuracy was obtained for simulating the daily evaporation data. The best results in the warm months were observed for Dezful and Abadan in August and for Ahvaz in May and in the M20. The error in these stations is 0.93 mm, 0.95 mm and 0.95 mm per day, respectively. The Nash- Sutcliffe coefficient (N.S) is more than 0.80 in all stations. In the cold months of the year, the lowest error (0.70 mm per day) was observed in December of Abadan, along with the Nash- Sutcliffe coefficient of 0.82, which indicates the high efficiency of the GEP model in data simulation. Comparing the results of the Gene Expression Programming model (GEP) with the studied time series models also showed that the GEP model has lower error and higher R2 and N.S coefficient in all months in Abadan. In Dezful, the Holt model provided better results than GEP in January and September. The error rate in these two months is 61% and 71% less than GEP model, respectively. In Ahvaz, the Exponential smoothing model in October and the Holt model in November estimated the daily pan evaporation with 48% and 68% less error, respectively.

Abstract Evapotranspiration plays an imperative role in management of regional water resources, climate change and agricultural production. In this study efficiency of some data-driven techniques, including support vector machine (SVM), artificial neural networks (ANN) and its hybrid with wavelet transform (WANN), multi linear regression (MLR) and decision tree (DT) for predicting Evapotranspiration rates at Scottsbluff Station in Nebraska have been monitored. For this purpose, 5 meteorological parameters utilized as inputs for the models. Daily meteorological information, data used in this study, were between 2005 - 2013 years to train and test the models. In order to implement each of the models 8 scenarios were considered according to combination of input parameters. For evaluate performance of the studied techniques, three different statistical indices were used which included root mean square error (RMSE), correlation coefficient (R) and Nash-Sutcliffe coefficient (NSE). In addition, Taylor charts were used to test similarity between observation and prediction data. The results showed that at the Scottsbluff station, WANN8 (is the eighth scenario for the WANN model) according to the root mean square error (RMSE), correlation coefficient (R) and Nash-Sutcliffe equal to 0.097, 0.999 and 0.999 performed better than ANN, SVM, MLR and DT. The SVM and ANN models also showed excellent accuracy, and the DT and MLR models performed worse than the other models despite their acceptable accuracy. As a conclusion, the results of the present study were proved that WANN provides reasonable procedures for modeling Scottsbluff at the Scottsbluff station.