Assessment of wavelet-SVR and wavelet-GP models in predicting the groundwater level using areal precipitation and consumption data

Abstract

ABSTRACT Assessing and predicting groundwater level (GWL) fluctuations using specific models provides valuable information for water resources management and consumption planning. In this study, areal monthly data of GWL precipitation and consumptions were used to predict the GWL using genetic programming (GP), wavelet-GP, support vector regression (SVR) and wavelet-SVR models over a period of 11 years. Appropriate time lags were determined using the autocorrelation function (ACF) and cross-correlation function (CCF). It was found that consumption with a two-month lag and precipitation with a one-month lag have the greatest effect on the GWL. Discrete wavelet transform (DWT) was used to decompose time series into low- and high-frequency components for wavelet-SVR and wavelet-GP models. Two types of input structures were used for modelling (single GWL data and GWL considering consumptions and precipitation data). The results showed that the wavelet-SVR model had the best performance (according to R, RMSE and NSE) compared to the other three models, and SVR had better performance than the GP.