Application of wavelet-neural network model for forecasting of groundwater level time series with non-stationary and nonlinear characteristics

Aquifer systems are often characterized by non-stationary and nonlinear features. Modelling of these systems and forecasting their future conditions requires identification of these fundamental features. Recently, wavelet analysis have been used widely in hydrological time series forecasting owing to its ability to decode aforementioned features. In this paper, a hybrid model based on coupling wavelet and artificial neural networks (WANN) that use sum of sub-series method, is tested for its ability to yield forecasts of groundwater level. The model results are compared with the results from artificial neural networks (ANN) and multi linear regression (MLR) models. The variables used to develop the models were monthly groundwater level at two piezometers and monthly total precipitation data recorded for 20 years in the Qom plain, Iran. Twelve-month-ahead prediction with the WANN model show that the error of this model is 30 and 23 percent less than ANN model and 37 and 51 percent less than MLR model for piezometers 1 and 2 respectively. The results show that precipitation has no significant effect on groundwater level variations of the two study piezometers; Although for the detail sub-series, use of precipitation improved the results.