دکتر جلال شیری

Pore water pressure, stress and settlement are the most important geotechnical parameters that must be constantly monitored during the construction of earth dams. Since measuring dam settlement directly at the time of dam construction requires cost and time, the development of artificial intelligence methods can be very effective. Most studies have been done in the field of modeling earth dams during construction with a numerical model; therefore the need for artificial intelligence modeling in this field seems to be necessary. Artificial intelligence models, including neural networks, are used for study and modeling many engineering sciences. Also, with the development of meta-heuristic algorithms, their combination with neural networks has become very wide-spreading due to more accurate results. So, The purpose of this study is to determine the most effective features in modeling settlement in the body of earthen dams at the time of construction as a case study (Kaboud-val dam) using the hybrid algorithm Dragonfly - artificial neural network in different points of the body of earthen dam at the time of construction. Therefore, in this research, new inputs in artificial intelligence modeling have been proposed for this purpose and their importance in different levels of installation has been investigated.

Kaboud-val Dam is located in Golestan province of Gorgan and around the city of Aliabad. This dam is homogeneous and has a filter and inclined drainage. In order to obtain the deformations of the body and the foundation of Kaboud-val dam, settlement plates have been installed in different sections of the body and its foundation during the construction. In this study, instrumental data related to the section of 19 Kaboud-val dam were used. Also, out of 17 pages, 4 pages named M1, M5, M9 and M13 (installed in the body and Kabudwal dam at levels 140, 152, 164 and 180 meter, respectively) have been used for modeling. By analyzing the data of section 19 pages, fill level (F), reservoir level (RL), dam construction time (T), fill rate (FR) and impounding rate (RV) for inlet and landing (P) on terms of (kp), was selected as the output of the hybrid model in the feature selection method. In this study, in order to select the best combination of input features in the artificial neural network, the dragonfly algorithm was used. Feature selection is a method of selecting a subset of related attributes (the best combination of them) that is relevant to a particular goal. The most important principle is to choose stable features and remove extra data. The combination of dragonfly algorithm with artificial neural network as DA-ANN is shown, Therefore, the dragonfly algorithm (DA) models a variety of different combinations of features with an artificial neural network and selects the best least error combination (RMSE) as the optimal artificial neural network model.

By performing a hybrid algorithm, sensitivity analysis and feature selection method, combining the four features on pages M1, M5,M9 and M13 with error values (RMSE) of 0.0023(kPa), 0.0024(kPa), 0.0026(kPa), respectively, and combining the three features on the M13 page with the value Error (RMSE) equal to 0.0035(kPa) was the best input combination. The three features of construction time, fill level and reservoir level as common features in all plates are the most effective features in modeling the settlement on selected plates. On plates mounted at higher levels, the modeling error increases, because during the test period and for plate M13 (with the highest mounting level), according to the statistical indices R^2, SI, NSE and NRMSE are equal to the values of 0.9998, 0.0062, 0.9998 and 0.0062 respectively, have poorer performance than other pages. The effect of reservoir level feature on the plates installed at higher levels due to the high sensitivity coefficient is more than other points and the fill level feature has the least effect on subsidence modeling.

The results are very important considering the cost of installing the measuring equipment and the significance of estimating the actual values in the future. The present study shows that the DA-ANN hybrid model is an important tool in predicting and selecting the best input combination for the intelligent model of the target variable of the settlement at the time of construction of earth dams. However, assessment of this model using the input data studied in different dams is necessary to ensure the application of these models in different conditions.

Various factors affect the location of the scour hole formed by the falling flow. Among these factors, we can mention the tailwater depth, the height of the falling flow, the velocity of the flow, the cross-section of the flow and the amount of air entering the flow. In the present study, the location of the formation of scour holes was investigated in the hydraulic laboratory of Tabriz University. Also, the ability of artificial neural networks (ANN) and tree models (M5P tree model) in estimating the location of scour holes downstream of dams was investigated using laboratory data and the results of these two models have been compared with the multivariate nonlinear regression method. The results showed that all three methods, the artificial neural network, the M5 tree model and regression method provide relatively accurate results in predicting the location of scour hole. RMSE value for ANN=1.75, M5=3.75 and Regression=3.89, but due to providing simple linear relationships by the M5 tree model, this method can be used as a practical method to determine the location of scour hole. The analysis of the M5 tree model showed that 4 equations with different linear equations model the pattern of changes in the location of scour hole. In addition, the analysis of the laboratory results showed that the regression equations presented in the present study compared to the common method (using projectile equations) have much less error when predicting the location of scour hole. Also, the laboratory results showed that the head passing through the structure is the most effective parameter in the formation of the scour hole.

Accurate prediction of pore water pressure in the body of earth dams during construction is one of the most important factors in managing the stability of earth dams. In this study, using three different evolutionary neural network models including multilayer perceptron neural network with genetic algorithm, particle swarm optimization and Imperialist Competitive algorithm for estimating the pore water pressure in the body of Kabudwal earth dam at the time of construction, has been studied. Five features including fill level, construction time, reservoir level, impounding rate and fill speed during the 4-year statistical period were selected as the input of the hybrid model in piezometer EP19.7. The composition of the inputs was obtained using the feature selection method and the hybrid water cycle algorithm -artificial neural network. By performing hybrid algorithm and sensitivity analysis and feature selection method, fill level, construction time, reservoir level and dewatering speed were selected as the top four inputs, because the combination of these four features with MSE value of 1.1587 had the least error. In this study, artificial neural network weights are optimized to increase efficiency using the above three meta-heuristic algorithms. In general, according to statistical indicators, the results indicate acceptable accuracy of all three hybrid models. In terms of priority, the ANN-GA hybrid model with the highest accuracy and minimum error and values of , RMSE and MAE are equal to 0.9773, 0.0457 and 0.0399, respectively, is first priority and ANN-PSO and ANN-ICA hybrid models were given the next priorities, respectively.

Accurate prediction of pore water pressure in the body of earth dams during construction with accurate methods is one of the most important components in managing the stability of earth dams. The main objective of this research is to develop hybrid models based on fuzzy neural inference systems and meta-heuristic optimization algorithms. In this regard, the fuzzy neural inference system and optimizing meta-heuristic algorithms including genetic algorithms (GA), particle swarm optimization algorithm (PSO), differential evolution algorithm (DE), ant colony optimization algorithm (ACOR), harmony search algorithm (HS), imperialist competitive algorithm (ICA), firefly algorithm (FA), and grey wolf optimizer algorithm (GWO) were used to improve training system. Three features including fill level, dam construction time, and reservoir level (dewatering) obtained from the dam instrumentation were selected as the inputs of hybrid models. The results showed that the hybrid model of the genetic algorithm in the test period had the best performance compared to other optimization algorithms with values of R2, RMSE, NRMSE, and MAE equal to 0.9540, 0.0866, 0.1232, and 0.0345, respectively. Also, ANFIS-GA, ANFIS-PSO, ANFIS-ICA, and ANFIS-HS hybrid algorithms performed better than ANFIS-GWO, ANFIS-FA, ANFIS-ACORE, and ANFIS-DE in improving ANFIS network training and predicting pore water pressure in the body earthen dams at the time of construction.

One of the basic measures in managing the stability of earth dams is to accurately estimate the amount of pore water pressure in the body of the dam during and after its construction. In this study, three different models of artificial neural network (ANN), adaptive neural-fuzzy inference system (ANFIS) and gene expression programming (GEP) to estimate the pore water pressure in the body of Kabudwal earthen dams at the time of construction have been studied and compared. Five features including fill level, construction time, reservoir level, impounding rate and fill speed have been used during the 4-year statistical period as input of models in 4 piezometers installed in the dam body. The first three features were the most effective inputs according to the cross-correlation function. In this study, the results obtained from artificial neural network (ANN) in two piezometers according to statistical indicators, provided more accurate answers than gene expression programming (GEP) and ANFIS, but in the other two piezometers it was the opposite. Also, ANFIS and GEP models provided more accurate answers in piezometers that had higher data scatter than ANN model. Finally, based on the GEP model, mathematical relationships between input features and output variables were extracted.

In this study, the ability of WCA-ANN hybrid algorithm to model the pore water pressure coefficient in the body of Kabudwal dam at the time of construction was investigated and the effective features were identified. Therefore, five features including fill level, time, reservoir level, dewatering rate and fill speed during the 4-year statistical period were selected as the input of the model. By running the hybrid algorithm and feature selection method, the two features of fill level and time at points RU19.1  and   RU19.2 have the greatest impact on modeling the pore water pressure coefficient. In addition to the above two features, in the points of the middle axis , the features of fill speed and reservoir level with error value (MSE) equal to 0.00006 and in points close to the dam reservoir, dewatering level and dewatering rate with error value equal to 0.00004 are effective in modeling the pore water pressure coefficient. The results showed that at points close to the dam axis, the fill level and at points farther from the middle axis construction time (with high sensitivity coefficient) was recognized as the most important features in modeling the pore water pressure coefficient with artificial intelligence models.

Accurate estimation of reference evapotranspiration (ETo) is an important issue in agricultural engineering, irrigation scheduling, water resources management, etc. Nowadays, the physics-based Penman-FAO-Monteith model is a standard model of ETo determining as well as calibrating other ETo equations. However, the need for large amount of hydro climatologic data has limited its application, especially in regions without sufficient necessary weather data. The present study aimed at assessing the performance accuracy of empirical temperature/radiation- based ETo models in humid regions and comparing the results with random forest (RF)-based models that used the same input variables of the applied empirical equations. The obtained results showed that the Random forest models gave the most accurate results when compared to the commonly used original and calibrated empirical equations (with similar input variables) with global MAE and SI values of 0.4 and 0.008. Nonetheless, the calibration procedure could not improve the performance accuracy of the equations in some cases.

Groundwater is an important source of exploitation in arid and semiarid regions. In the Azarshahr plain, due to the reduction in annual precipitation and lack of surface runoff, groundwater resources play an important role in supplying agricultural, drinking, industrial and domestic water requirements. The aim of this study was to evaluate the quality of groundwater resources in Azarshahr plain for agricultural irrigation and to analyze the EC and SAR variations by using geostatistical methods. So, for spatial mapping of SO4, Cl, TDS, EC and SAR, the observational data from 47 wells in this region, which were measured during the years 2003-2017, were utilized. In order to select the best interpolation method, Kriging approach with Gaussian variogram and IDW methods were used. The obtained results revealed that the Kriging with Gaussian variogram showed the lowest RMSE and the highest NS values. The results of the groundwater quality classification (using Wilcox diagram) and EC/ SAR zoning showed that the groundwater quality in the Azarshahr plain is very saline (C4S1) in most of the cases, which might be linked to the higher amounts of withdrawals from the aquifer.

In order to determine and analyze the head losses caused by the combined effect of entrance resistance and approach flow convergence to drains, some known models were studied. The field data of an existing sub-surface drainage system were applied for evaluating the models and selecting the best one. The hodograph solution was employed to compute the water head, directly above the drain, while the Oosterbaan model was used to validate the hodograph results. Total head loss between two drains, was determined using Ernst equation. The results showed that increasing total head loss caused reduction of hydraulic head in approach flow region in a non-linear manner. Application of Youngs and Hoffman equations as well as EnDrain program based on energy and Darcy equations could simulate drain discharge to some extent. Hoffman equation gave some promising results among the others in term of R (0.928), RMSE (0.0007) and SI (1.09).