Izvorni znanstveni članak

https://doi.org/10.17559/TV-20260625003684

Leakage-Free Depth-to-Groundwater Forecasting and Monitoring-Well Vulnerability Screening Framework for Incomplete Groundwater Monitoring Networks

Xuanru He ; School of Electronic Engineering, North China University of Water Resources and Electric Power, No. 136Jinshui East Road, Zhengzhou City, Henan Province, China
Yiyang Zhao ; School of Geosciences and Engineering, North China University of Water Resources and Electric Power, No. 136 Jinshui East Road, Zhengzhou City, Henan Province, China
Peide Lin ; School of Water Conservancy, North China University of Water Resources and Electric Power, No. 136 Jinshui East Road, Zhengzhou City, Henan Province, China
Zixuan , Xu ; School of Geosciences and Engineering, North China University of Water Resources and Electric Power, No. 136 Jinshui East Road, Zhengzhou City, Henan Province, China
Jiongyi Li ; School of Information Engineering, North China University of Water Resources and Electric Power, No. 136 Jinshui East Road, Zhengzhou City, Henan Province, China
Kai Zhao ; School of Information Engineering, North China University of Water Resources and Electric Power, No. 136 Jinshui East Road, Zhengzhou City, Henan Province, China *

* Dopisni autor.

Sažetak

Long-term groundwater monitoring networks are often affected by interruptions in manual measurements, sensor or equipment failures, and monitoring-well heterogeneity. These factors generate point-wise random missingness, block-wise missingness, seasonal missingness, and mixed missingness, which cause difficulty for average forecasting errors to identify which monitoring wells are more vulnerable under incomplete observations or how limited maintenance resources should be allocated. To reveal the relationship among depth-to-groundwater forecasting errors, missing-data sensitivity, and monitoring-well maintenance priority, a leakage-free multi-scenario forecasting framework and a station vulnerability score (SVS) were proposed, where station denotes an individual groundwater monitoring well. A chronological train-validation-test split was used, four missing-data scenarios were constructed, a residual long short-term memory model with station embeddings (Residual-SE-LSTM) was developed, and comparisons were made with Persistence, Ridge, Random Forest, and a histogram-based gradient boosting (HistGradientBoosting) baseline. Block-wise and mixed missingness were used to simulate contiguous observation gaps and the combined effects of multiple missing-data sources, while budget constraints were used to represent limited maintenance resources for groundwater monitoring networks. The stability and screening value of SVS were further evaluated through bootstrap resampling, weight sensitivity analysis, forward validation, and budget-constrained experiments. Results show that Residual-SE-LSTM achieves an average root mean square error of 1.547 m and an average Nash-Sutcliffe efficiency of 0.961 across 36 formal forecasting scenarios. The early-SVS achieves an area under the receiver operating characteristic curve of 0.829 for identifying the later-period worst 10% error wells and covers 58.82% of the later-period SVS-critical wells under a 10% maintenance budget. These findings provide quantitative support for forecasting-reliability assessment, vulnerable monitoring-well identification, and maintenance prioritization in groundwater monitoring networks with missing data.

Ključne riječi

depth-to-groundwater forecasting; leakage-free evaluation; maintenance prioritization; missing data; monitoring-well vulnerability; Residual-SE-LSTM

Hrčak ID:

350443

URI

https://hrcak.srce.hr/350443

Datum izdavanja:

31.8.2026.

Posjeta: 0 *