Scholarlink

<p>Karst springs provide drinking water for approximately 10% of the world’s population. Reliable prediction of spring discharge and water quality is therefore crucial for sustainable water management. Although machine learning (ML) models have shown great potential for forecasting hydrological variables in recent years, the understanding of the underlying processes remains limited. The aim of this study was to increase the transparency of ML models through an attribution analysis, which examines the contribution of local environmental factors to model predictions. The study focused on karst springs draining the Hochschwab Massif in the Eastern Alps. Several ML models were employed to forecast both spring discharge and water quality—described by the spectral absorption coefficient at 254 nm (UV254)—with a&#xa0;prediction horizon of up to four days. The application of the Deep-SHAP method allowed for the identification of pronounced seasonal variations in model attributions, with the strongest changes observed for snow depth, followed by physicochemical variables such as electrical conductivity, as well as other meteorological factors. Among the tested models, the Transformer model showed the best overall predictive performance. Model uncertainty, quantified using the Deep Ensemble method, was higher in spring and summer, with both model errors and uncertainties increasing with the variability of the target variables. To evaluate the practical applicability of the models for selective water abstraction, UV254 predictions were classified based on threshold exceedances, achieving high classification accuracies (&gt; 95% for 1‑day and &gt; 90% for 2‑day forecasts). The combination of the Deep-SHAP and Deep Ensemble methods enhances the transparency of data-driven models and provides valuable insights that can support decision-making in the management of drinking water from karst systems. This article is an adapted and slightly shortened version of the peer-review article by Pölz et al., “Improving transparency in karst spring discharge and water quality forecasts using interpretable machine learning models in the Eastern Alps” which was published in <CitationRef CitationID="CR34">2026</CitationRef> in the <i>Journal of Hydrology: Regional Studies, Vol</i>. 64, article number 103147. This topic was realized as part of the Vienna Water Resource Systems 2020+ research collaboration (VIWA&#xa0;20+) and presented at the Mid-Term Symposium 2025.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

KI-Modellierung für das Quellmanagement: Vorhersage von Karstquellschüttung und Wasserqualität mittels interpretierbarer Machine-Learning-Modelle

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月20日

查看全文
  • Anna Pölz,
  • Alfred Paul Blaschke,
  • Katalin Demeter,
  • Günter Blöschl,
  • Margaret E. Stevenson,
  • Helene Bauer,
  • Liping Pang,
  • René Mayer,
  • Andreas H. Farnleitner,
  • Julia Derx

摘要

Karst springs provide drinking water for approximately 10% of the world’s population. Reliable prediction of spring discharge and water quality is therefore crucial for sustainable water management. Although machine learning (ML) models have shown great potential for forecasting hydrological variables in recent years, the understanding of the underlying processes remains limited. The aim of this study was to increase the transparency of ML models through an attribution analysis, which examines the contribution of local environmental factors to model predictions. The study focused on karst springs draining the Hochschwab Massif in the Eastern Alps. Several ML models were employed to forecast both spring discharge and water quality—described by the spectral absorption coefficient at 254 nm (UV254)—with a prediction horizon of up to four days. The application of the Deep-SHAP method allowed for the identification of pronounced seasonal variations in model attributions, with the strongest changes observed for snow depth, followed by physicochemical variables such as electrical conductivity, as well as other meteorological factors. Among the tested models, the Transformer model showed the best overall predictive performance. Model uncertainty, quantified using the Deep Ensemble method, was higher in spring and summer, with both model errors and uncertainties increasing with the variability of the target variables. To evaluate the practical applicability of the models for selective water abstraction, UV254 predictions were classified based on threshold exceedances, achieving high classification accuracies (> 95% for 1‑day and > 90% for 2‑day forecasts). The combination of the Deep-SHAP and Deep Ensemble methods enhances the transparency of data-driven models and provides valuable insights that can support decision-making in the management of drinking water from karst systems. This article is an adapted and slightly shortened version of the peer-review article by Pölz et al., “Improving transparency in karst spring discharge and water quality forecasts using interpretable machine learning models in the Eastern Alps” which was published in 2026 in the Journal of Hydrology: Regional Studies, Vol. 64, article number 103147. This topic was realized as part of the Vienna Water Resource Systems 2020+ research collaboration (VIWA 20+) and presented at the Mid-Term Symposium 2025.