Groundwater Level Forecasting using Radial Basis Function and Generalized Regression Neural Networks

Abstract

Forecasting of groundwater levels is very much useful for efficient planning in integrated management of groundwater and surface water resources in a basin. Accurate and reliable groundwater level forecasting models can help ensuring the sustainable use of a watershed’s aquifer for both urban and rural water supply. The present work investigates the potential of two Neural networks, such as Radial Basis Function Neural Networks (RBFNN) and Generalized Regression Neural Networks (GRNN) in comparison to regular ANN models like Feed Forward Back Propagation (FFBP) and Non-Linear Regression Model (NARX) for modeling in Ground water level (GWL) forecasting in a coastal aquifer at western Ghats of India. Total 24 wells (both shallow and deep) located within the study area (microwatershed of Pavanje river basin) were selected covering around 40sqkm. Here, two different dataset such as weekly Time series GWL and Meteorological variables those recorded during the study period (2004-2011) were used in the analysis. Various performance indices such as Root Mean Squared Error (RMSE), Coefficient of Correlation (CC) and Coefficient of Efficiency (CE) were used as evaluation criteria to assess the performance of the developed models. At the first stage, the potential and applicability of RBF for forecasting groundwater level are investigated. Weekly time series groundwater level data upto four lagged data has been used as various input scenario where predicted output are one and two week leadtime GWL. The analysis has been carried out separately for three representative open wells. For all the three well stations, higher accuracy and consistent forecasting performance for RBF network model was obtained compared to FFBP network model. After confirming the suitability of RBF in GWL forecasting and with better accuracy over FFBP, the work has been extended further to consolidate the applicability of RBF in multistep leadtime forecasting upto six week ahead. In this study, six representative wells are covered for development of RBF models for six different input combinations using lagged time series data. Outputs are the predicted GWL upto six week. RBF models are developed for every well station and results are compared with Non linear regression model (NARX). It has been observed that for allGroundwater level Forecasting using Radial Basis Function and Generalized Regression Neural Networks, Ph.D Thesis, 2012, NITK, Surathkal, India viii the six well station, the higher and consistent forecasting performance by RBF network model in multi step week lead which consolidates the forecasting capability of RBF. The NARX model result shows poor performance. In the third stage, to examine the potential and applicability of GRNN in GWL forecasting, various GRNN models has been developed by considering the advantage of S-summation and D-summation layers for different input combinations using time series data. Weekly time series groundwater level data upto four lagged data has been used as inputs where predicted outputs are one week leadtime GWL. The analysis has been carried out separately for three representative open wells. GRNN models were developed for every well and best model results were compared with best RBF and FFBP with LM training algorithm models. The RBF and GRNN models are almost performed similarly in GWL forecasting with higher accuracy in all the representative well station. The poor performance of FFBP-LM model is also satisfactory but found inferior than both GRNN and RBF. After confirming the potential and applicability of GRNN and RBF in time series GWL forecasting with similar capability, the robustness, adaptability and flexibility characteristics of these two techniques are further investigated for suitability with cause and effect relationship. Here various meteorological parameters are used as causable variable and the GWL is used as output effect .Only GRNN models are developed in the present study as RBF was found with similar predicting performance in previous studies. Five various input combinations are used to obtain best results as one step leadtime output for three representative wells. In this case also, GRNN model is predicting groundwater level with higher accuracy and with satisfactory results. The GRNN model performance is compared to general ANN (FFBP) model and found outperforming FFBP performance. The result of the study indicates the potential and suitability of RBFNN and GRNN modeling in GWL forecasting for multistep leadtime data. The performance of RBFNN and GRNN were found almost equally good. Although accuracy of forecasted GWL generally decreases with the increase of leadtime, the GWL forecast were obtained within acceptable accuracy for both the models.