Prediction of fluctuations in the underground water level of Sanghar Plain using machine learning methods

The fluctuation of underground water level is one of the important criteria required for decision-making in many water resources exploitation models. The lack of reliable and complete data is one of the most important challenges in analyzing the decline and predictions of the underground water level in water management. In recent years, the use of different numerical models has been noticed as a reliable solution. These models are able to estimate based on extensive statistics and information and based on various land maps and measurements such as pumping tests, geophysics, soil and land use maps, topography and slope data, different boundary conditions and using complex equations. The level of underground water in any region.The studied area is Sanghar plain in the west of Iran, located at a distance of 100 km northwest of Kermanshah city (Figure (1)). Sanghar plain is one of the fertile plains in Kermanshah province, whose needs are provided by two systems of surface water and underground water. Part of the water needed in the plain is provided by Suleimanshah Dam (Shahda) and the rest is provided by 278 deep wells dug in the south and west of the plain. 

In the present research, first, by using available statistics and information and maps, the fluctuations of the underground water level of Sanghar Plain were simulated by the GMS model, and the accuracy of the model was evaluated in two stages of calibration and validation. Then, due to the need for much less data volume in machine learning methods, GWO-ANN and PSO-ANN hybrid methods and LSTM and SAELM models were used. Based on the general direction of the underground water flow in the entire Sanghar plain, the grid direction was considered to be 250x250 meters in the north direction. Therefore, the model network was built with 2596 cells (44 rows and 59 columns) with 250 meters intervals, which included 908 active cells and 1688 inactive cells. In this study, the general load boundary package was used to simulate the entry and exit borders of Sangar plain. In this package, the inlet or outlet flow is affected by the hydraulic gradient at the boundary and the conductance of the boundary cell. Using the prepared geophysical sections and the data log of the wells, a rock map of the plain was prepared. Also, the DEM map of the plain was used to determine the upper limits of the layer in the groundwater model. In the GMS model, the WELL package was used to simulate exploitation wells in Dasht Sangar (278 wells) and well cells were identified. The feeding of the plain is one of the important parameters in the groundwater model. Usually, due to the different characteristics of soil, geology, vegetation, rainfall intensity and the slope of the land, the amount of groundwater recharge is different in different places. In the GMS model, the RCH package is used to consider the feeding. The zoning method was used to estimate the hydrodynamic parameters of the aquifer. The zoning of the area for hydraulic guidance and special drainage was done based on the drilling logs of observational, exploratory and piezometric wells, as well as geophysical sections prepared from the area. According to the type of soil and sediments of each zone, the initial values of hydraulic conductivity and specific drainage were estimated. Finally, after performing the calibration process, for each zone, the optimized value of hydraulic conductivity and specific drainage was taken into account. In the underground water simulation section, after the calibration and validation tests of the model in two permanent and non-permanent modes and ensuring its accuracy, the final zoning of the main parameters of the model, i.e. hydraulic conductivity and specific drainage, was prepared so that the model can predict the changes in the underground water level for 6 years. Simulate consecutively. Because all the required information was available for 6 years (October 2019 to September 2015). 

The results showed that the output of the SAELM model had the best fit with the observational data with a correlation coefficient equal to 0.97, and it also had the best and closest distribution of points around the 45 degree line, and in this sense, it is considered the most accurate model. Therefore, to predict the level of underground water in the whole plain, instead of using the complex GMS model with a very large volume of data and also a very time-consuming calibration and validation process, SAELM model can be used with confidence.  

This approach greatly helps the researchers of the underground water sector to predict the changes of the underground water level in dry and wet years without using numerical models with a complex and time-consuming structure using artificial intelligence with high accuracy