Machine-Learning Methods for Water Table Depth Prediction in Seasonal Freezing-Thawing Areas

Abstract

Long-term and accurate predictions of regional groundwater hydrology are important for maintaining environmental sustainability in arid agricultural areas that experience seasonal freezing and thawing where serious water-saving measurements are used. In this study, we firstly developed a machine-learning method by integrating a multivariate time series controlled auto-regressive method and the ridge regression method (CAR-RR) for water table depth modeling. We applied and evaluated this model in the Hetao Irrigation District, located in northwest China where the freezing-thawing period is 5 months long. To train and validate the model, we used monthly data of water diversion, precipitation, evaporation, and drainage from 1995 to 2013. The CAR-RR model yielded more accurate results than the support vector regression (SVR) and multiple linear regression (MLR) models did in the validation period. To extend the model applicability during freezing-thawing periods, we included additional temperature information. We compared results obtained using temperature only during the freezing-thawing period with results obtained without temperature, which showed that the input data of the temperature during the freezing-thawing period significantly improved the model accuracy. To resolve the problem of capturing the peaks and troughs of CAR-RR, we further developed an integrated CAR-SVR model to consider the nonlinearity. The optimal model (CAR-SVR) was then used to predict the water table depth under future water-saving measurements. It demonstrated that water diversion was the most important factor affecting the water table depth. A water table depth with less than 3.64 billion m³ water diversion will result in risks of environment problems.