Comparison of ARIMA Time Series, Multi Linear Regression and Artificial Neural Network Model for Prediction of the Variations of Groundwater Level

Prediction of groundwater level fluctuations is an essential for planning in arid and semi-arid regions. In this study, artificial neural network models, ARIMA time series and multivariate linear regression models were used to predict groundwater fluctuations of two piezometers located in Kerman plain. In order to achieve this goal, the depth of groundwater of the monthly piezometers was used during the years 2002-2013. The results of studying different models of ARIMA model showed that the ARIMA (0,1,1) and (2,0,2) for south Baghin piezometer and ARIMA (1,1,1) and (2,0,0) for Airport areas piezometer are the best-fit time series model with the data. In the model of MLP and RBF artificial neural network, MLP with 2 and 4 layers of hidden and RBF with 8 and 10 hidden layers for southern Baghin piezometers and the airport areas have the best fit with the data. In multivariate linear regression modeling, for each of the two piezometers, the best correlations of the multivariable linear regression model show that the multivariate linear regression relationship of groundwater depth of the current month is a function of groundwater depth of one month prior; in other words, the depth of water the ground water level has the highest dependence on groundwater depth of its prior month. The results showed that prediction of groundwater depth by multivariate linear regression model is better than Neural Network and ARIMA model.