پيش بيني زماني ومكاني سطح آب زيرزميني محدوده متروي شهر تبريز توسط روش كريجينگ عصبي

Forecasting Spatiotemporal Water Levels by Neural Kriging Metbod in Tabriz City Underground Area

Introduction Groundwater level forecasting is one of the most important factors in hydrogeological studies. Artificial Neural Network (ANN) models are among the black box models much suitable for nonlinear systems. The ability to identify a relationship from given patterns make it possible for ANNs to solve complex hydrologic problems. Concepts and applications of the ANN models in hydrology have been discussed by the ASCE (2000) and Asghari Moghaddam et a1. (2008). Neural networks have also been previously applied successfully to temporal prediction of groundwater levels in alluvials (Daliakopoulos et aI., 2005). Geostatistics is another powerful interpolating approach widely utilized in spatial estimation of hydrological variables (Goovaerts, 2000) such as groundwater level mapping (Hoeksema et al., 1989). Rizzo and Dougherty (1994) introduced the idea of Neural-kriging for characterization of aquifer properties. Objective, In some parts of the Tabriz underground construction project the groundwater levels are high which show the necessity of groundwater modeling. Due to the high complexity of Tabriz aquifer, the combination of ANN and geostatistic models for spatiotemporal prediction of groundwater levels are used in the study area. Results were applied as inputs of the geostatistical model for spatial forecasting ofgroundwater levels. Methodology Several neural networks are evaluated in order to reach the required efficiency and to find the suitable technique for presenting ANN models. Two popular neural network models are the Feedforward Neural Network (FNN) and the Recurrent Neural Network (RNN). Different algorithms can be used for training these networks. In the present research, gradient descent with momentum and adaptive learning rate backpropagation (GDX), Levenberg-Marquardt (LM), and Bayesian Regularization (BR) are used. Some studies provided details of the used networks and algorithms in ANN modeling (e.g.. Coulibaly et al., 2001a). Several aspects of the architecture (structure) of neural networks that focus on the prediction of variables associated with hydrology are covered by Maier and Dandy (2000). Their suggestions were followed for temporal water level forecasting of Tabriz aquifer and the structures of the network were determined by trial and error. The size of the input and hidden layers of the network have been determined considering the prediction target and the output layer has just a single node; predicted water level. The activation function of the hidden layer was set to a hyperbolic tangent sigmoid function as this proved by trial and error to be the best in depicting the nonlinearity of the modeled natural system. among a set of other options (linear and log sigmoid). Finally, the number of the hidden layer nodes and the training epoch number were optimized in terms of obtaining precise and accurate output.Presenting a hybrid neural-geosratistics (NG) modcl, the external 24 monthly temporal data of ANN models were used as the input data of the geostaristic modeL The model can use different methods depending on the subject or the problem. In order to obtain the best method of the geostatistical approach, different methods were examined; I) Nearest Neighbor (N.N.), 2) Kriging, 3) Polynomial Regression (P.R.), 4) Radial Basis Function (R.B.F.), 4) Moving Average (M.A.), 5) Inverse Distance to a Power (I.D.P.), 6) Minimum Curvature (M.C.), 7) Modified Sheperds Method (M.S.M.) and 8) Local Polynomial (L.P.). These methods were described in detail by Isaaks and Srivastava (1989). -1 391836 -1 181029 R.esiduals , 602000 4214950 12 136343 Table L Residuals for predicted groundwater level in testing step, For achieving the neural-geostatistical spatiotemporal forecasting model the external 24 monthly temporal data of ANN models were utilized as input data in the geostatistic model for spatial connections. In order to obtain the best method of the geostatistical approach, different methods (mentioned in methodology) were examined. In order to obtain the best geostatistical method, the first month of predicted data by ANN model for considered piezometers were fed into the eight methods as input data sets. The N.N method presented the best method connection among the forecasting data. This method was then utilized for making connection among other 24 months of forecasting, so that for each month. forecasted data produced a neural-geostatistical model. Similarly monthly predicted contours can also be presented (for I th to 24ʹ" month). The efficiency of the hybrid neuralgeostatistic model were tested by the four piezometers not utilized in model calibration. The results of this testing are presented in Table I. values by FNN-LM Model, A) Training step, B) Test step, Conclusion ANN and geostatistic models each have a high ability in spatial and temporal prediction of hydrogeological phenomena. The combination of these models can result in an improved tool. In this research, a combination of the models were utilized as a hybrid spatiotemporal model of groundwater level prediction in multilayer aquifer in Tabriz city area. The obtained results represented suitable spatiotemporal prediction of groundwater levels in the study area. Methodology of this study can solve most of the existing problems in spatiotemporal groundwater level prediction in aquifer spatially multilayer aquifers. Furthermore, results of this research could prevent spending high costs for borehole drillings in Tabriz underground construction.