Integrated preprocessing techniques with linear stochastic approaches in groundwater level forecasting

Abstract

Accurate modeling of groundwater level (GWL) is a critical and challenging issue in water resources management. The GWL fluctuations rely on many nonlinear hydrological variables and uncertain factors. Therefore, it is important to use an approach that can reduce the parameters involved in the modeling process and minimize the associated errors. This study presents a novel approach for time series structural analysis, multi-step preprocessing, and GWL modeling. In this study, we identified the time series deterministic and stochastic terms by employing a one-, two-, and three-step preprocessing techniques (a combination of trend analysis, standardization, spectral analysis, differencing, and normalization techniques). The application of this approach is tested on the GWL dataset of the Kermanshah plains located in the northwest region of Iran, using monthly observations of 60 piezometric stations from September 1991 to August 2017. By removing the dominant nonstationary factors of the GWL data, a linear model with one autoregressive and one seasonal moving average parameter, detrending, and consecutive non-seasonal and seasonal differencing were created. The quantitative assessment of this model indicates the high performance in GWL forecasting with the coefficient of determination (R2) 0.94, scatter index (SI) 0.0004, mean absolute percentage error (MAPE) 0.0003, root mean squared relative error (RMSRE) 0.0004, and corrected Akaike s information criterion (AICc) 151. Moreover, the uncertainty and accuracy of the proposed linear-based method are compared with two conventional nonlinear methods, including multilayer perceptron artificial neural network (MLP-ANN) and adaptive neuro-fuzzy inference systems (ANFIS). The uncertainty of the proposed method in this study was\u2009±\u20090.105 compared to\u2009±\u20090.114 and\u2009±\u20090.126 for the best results of the ANN and the ANFIS models, respectively.