Stochastic analysis of groundwater level fluctuations at wells supplying freshwater for Izmir city

Izmir, the third largest city of Turkey, has an annual municipal fresh water demand of around 200 million m3 /year. Over half of its fresh water requirements is supplied from wells drilled in the Göksu and Sarıkız aquifers. Therefore, modelling of temporal behaviour of these water resources bears a vital importance. In this study, deterministic trends in monthly groundwater levels recorded at Göksu 16859-A and Sarıkız Rasat-1 wells are identified by various trend tests and removed. The non-seasonal linear stationary time series models, such as AR(1), AR(2) and ARMA(1,1), are fitted for the detrended fully standardized water levels. The non-stationary linear ARIMA(p,1,q) models are fitted for the first order differenced and fully standardized sample series. It is found that the first order differenced and fully standardized water levels in both wells can successfully be represented by a random walk model, ARIMA(0,1,0), or by a non-stationary integrated moving average, IMA(1,1), model. Secondly, the single input-single output transfer function models, such as ARMAX(1,1,1,0), ARX(1,1,0), ARX(2,1,0) are developed to explore the effect of rainfall input on the water levels. Five different artificial neural network (ANN) models are also developed to compare the estimation performances of those nonstationary linear stochastic models. It has been found that forecasting performance of the ANN models that incorporate the preceding water levels, precipitations and water withdrawals in the same month were better, and they were able to represent the nonlinear evolution mechanism of the actual water level processes. Keywords: Groundwater level fluctuations, stationary and non-stationary time series modeling, transfer function models, artificial neural networks (ANN)