Support Vector Regression Approach for Predicting Groundwater Levels under Variable Pumping and Infiltration Conditions

Abstract

Regression problems in environmental engineering can be tackled in principle by fundamentally different approaches, e.g., physically based numerical modeling or methods of machine learning. Appliance of the machine learning method Support Vector Machines (SVM) for regression is called Support Vector Regression (SVR). The feasibility of SVR for predicting groundwater levels in complex groundwater systems under variable pumping and infiltration conditions is demonstrated in a representative study area of groundwater management. Real-world data were used to train SVR models to predict transient groundwater levels in response to changing pumping and infiltration conditions. The SVR models were then validated with twelve sequential months. The prognoses of one year in monthly periods were compared against measured groundwater levels. Although the experiments are still in an early state, the best SVR models so far already achieve in groundwater level prognosis of twelve months an average monthly deviation of about 0,029m between the SVR predicted and the measured water level. While different SVR models can retain unlike qualities in terms of diverse scenarios and the climate scenario that serves as input data for the prediction horizon is uncertain within a certain scope, more than one SVR model and climate scenarios may be combined in an ensemble fashion. To put it in a nutshell, it is to say that the deployment of SVR technology in groundwater prediction holds the fundamental potential to improve management strategies and sound decision-making for hydro geological problems.