Simulation of Malayer Plain Groundwater Level Based on Weather Data Using Artificial Neural Network

Extended Abstract

Introduction
Prediction of groundwater level is necessary for supply management and utilization of water. Groundwater Fluctuations is influenced by many variables. One of the most appropriate methods, the study of groundwater, is using computer models. Thus, understanding of ground water variation mechanism and its prediction is necessary for awareness of available groundwater. In this regard, artificial neural network models, due to lack of understanding of the physical nature of the problem in modeling a nonlinear system are capable. Studies in Iran show that, in spite of the importance of groundwater level prediction and intelligent neural network capabilities, few studies regarding water table simulations using artificial neural networks has been done so far.
According to the necessities expressed purpose of this study is establish the relation between of groundwater level fluctuations associated with effective artificial neural network, in order to simulate and quantify associated water table fluctuations as an independent variable with various meteorological factors and hydro geology of Malayer plain.
Methodology
Malayer plain groundwater table located in Hamedan province was selected to evaluate the efficiency of artificial neural network for predicting water table changes. The collected data were based on similar studies and the effects of data on water table. They were used as inputs for artificial neural network on four structures. First data structure was including average maximum air temperature, minimum air temperature, maximum relative humidity, minimum relative humidity and evaporation monthly scale and height of water table of last month. In the second structure, water table data for every month from 1, 2, 3 and 4 previous months were used as inputs. Average of water table of mount and average of monthly water table plus the input of second structure were used as input for third structure. The inputs of fourth structure were mean of water table, data of last month and monthly meteorological data. Different patterns of artificial neural network were used on the basis of the four structure using Neurosolution software to estimate the water table.

Results and Discussion
The results showed that using water table parameters of previous years had the higher precision than other parameters on the prediction of water table. In other words, the results of the second and third data structures in which the input data kind was the same as output one better than two other structures were evaluated. In total, third structure topology 1-4-4-6 shows 6 neurons in the network input and middle layer 4 neurons is a suitable structure with 1.9 percent error compared with the actual values. The research showed that hydrograph of the plain during 1995 to 2006 was descending with the average of 1.2 meter annual loss and in total study period 14.5 m loss has experienced. Implementation of optimal neural network model consists of 6 neurons in the first layer of information; 1.18 m loss water table can be estimated. In other words, neural network with 1.9 percent average error rate was able to estimate withdraw from the table during studies 313 million m3 versus 319 million m3 of the actual harvested. Root Mean Square Error (RMSE) in the neural network optimal model 1-4-4-6 pattern based on Levenberg Marquet rule learning and the sigmoid function against table level actual changes was obtained 0.44 m with coefficient of determination 0.99 respectively. Error values resulting from the implementation of the proposed neural network model compared to other structures and similar studies have been done a very good accuracy.

Conclusion
Totally, it could be inferred that such variables entered the model structure based on structure of the number 3, the model and prediction water table changes is successful.
Comparison of the optimum structure of neural network study with Izadi et al (2007) and Esmail verki et al (2004) in predicting groundwater levels, confirming the good performance of the network in this study with six input parameters to 15 parameters in the report and Izadi et al (2007) (R2 = 0.937) and 10 parameters in the study of sciatic Ismail et al (2004) (R2 =0.90) is. Meanwhile the number of input neurons in neural network under the terms of the speed and accuracy is more preferable (Abyaneh Zare et al, 2010).

The suitable model and carefully downtrend governing table, using artificial neural networks for decision and management plain, as appropriate tool with speed and accuracy in predicting groundwater level in Malayer plain is recommended.