Reliable long-term groundwater level (GWL) prediction is essential to assess the availability of resources and the risk to drinking water supply in changing climatic and socio-economic conditions, especially in areas with water deficits. The modern approach in this area involves the use of machine learning methods. However, the greatest challenge in these methods lies in the optimization of input selection. The presented research concerns the selection of the best combination of predictors using the Hellwig method. It served as a preprocessing technique before GWL prediction using support vector regression (SVR) and multilayer perceptron (MLP) for three wells in the Greater Poland Province, where the largest water deficits occur, in the period 1975–2014. The results of this method were compared with those of the regression method, general regression model. For the case study under investigation, the Hellwig method provided the best set of predictors consisted of GWL at lags of −1 and −2 months, precipitation from the current month, and delayed from −1 to −6 months, and past temperature at −1, −3, −4 and −6 months. Such input led to a model accuracy of 0.003–0.022 for a mean squared error and r² of >0.8. The results obtained with SVR were slightly better than those with MLP. Moreover, every well required an individual set of predictors, and additional meteorological inputs improved the models’ performance.

可靠的长期地下水位 (GWL) 预测对于评估资源的可用性以及在不断变化的气候和社会经济条件下对饮用水供应的风险至关重要，尤其是在缺水地区。该领域的现代方法涉及使用机器学习方法。然而，这些方法中最大的挑战在于输入选择的优化。所提出的研究涉及使用 Hellwig 方法选择预测变量的最佳组合。它作为 GWL 预测之前的预处理技术，使用支持向量回归 (SVR) 和多层感知器 (MLP) 对 1975 年至 2014 年期间发生最大缺水的大波兰省的三口井进行。将该方法的结果与回归方法的结果进行比较，一般回归模型。对于正在调查的案例研究，Hellwig 方法提供了一组最佳预测因子，包括滞后 -1 和 -2 个月的 GWL、当月的降水量以及从 -1 到 -6 个月的延迟，以及过去的温度 - 1、-3、-4 和 -6 个月。这样的输入导致均方误差的模型精度为 0.003–0.022，并且r ² > 0.8。使用 SVR 获得的结果略好于使用 MLP 获得的结果。此外，每口井都需要一组单独的预测器，额外的气象输入提高了模型的性能。