Scholarlink

<p>Accurate simulation of groundwater level changes is important for sustainable water management, especially in aquifers where monitoring data are incomplete. Missing data are common in such datasets and can affect model performance. This study proposes a two-step approach that aims to fill data gaps and improve predictive performance. The framework is applied to monthly groundwater levels from 15 observation wells in the Ajab Shir Plain, northwestern Iran, covering the years 2006–2022. The data are divided into training (70%) and test (30%) sets. To reduce the complexity of modeling, predicting with three lags was evaluated in two scenarios. In the first scenario, data with missing values are removed or completed with the arithmetic average, while in the second case, the missing data are reconstructed with the Deep Missing Value Imputation (DeepMVI) model. Then investigated how this reconstruction affects prediction using four models. Models used for prediction are a Patch-based Time Series Transformer (PatchTST), a Gated Recurrent Units with Long Short-Term Memory (GRU-LSTM) hybrid model, an Artificial Neural Networks (ANN), and support vector regression (SVR). The results showed that the models trained on the reconstructed data performed better than the models trained on the dataset in which the missing observations were removed. Under the reconstruction scenario, PatchTST had a coefficient of determination (R²) of 0.961, a root mean square error (RMSE) of 3.203&#xa0;cm, and a Nash-Sutcliffe efficiency (NSE) of 0.960, outperforming the other four models. This suggests that even in areas with limited monitoring, advanced imputation methods can be useful in improving the reliability of predictions.</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

Enhancing groundwater level prediction through DeepMVI‑Aided interpolation and transformer‑based modeling

  • 期刊论文

  • 发表日期：

    2026年5月26日

查看全文
  • Aynaz Vafaei,
  • Osama Ragab Ibrahim,
  • Erfan Abdi,
  • Esmaeil Asadi

摘要

Accurate simulation of groundwater level changes is important for sustainable water management, especially in aquifers where monitoring data are incomplete. Missing data are common in such datasets and can affect model performance. This study proposes a two-step approach that aims to fill data gaps and improve predictive performance. The framework is applied to monthly groundwater levels from 15 observation wells in the Ajab Shir Plain, northwestern Iran, covering the years 2006–2022. The data are divided into training (70%) and test (30%) sets. To reduce the complexity of modeling, predicting with three lags was evaluated in two scenarios. In the first scenario, data with missing values are removed or completed with the arithmetic average, while in the second case, the missing data are reconstructed with the Deep Missing Value Imputation (DeepMVI) model. Then investigated how this reconstruction affects prediction using four models. Models used for prediction are a Patch-based Time Series Transformer (PatchTST), a Gated Recurrent Units with Long Short-Term Memory (GRU-LSTM) hybrid model, an Artificial Neural Networks (ANN), and support vector regression (SVR). The results showed that the models trained on the reconstructed data performed better than the models trained on the dataset in which the missing observations were removed. Under the reconstruction scenario, PatchTST had a coefficient of determination (R²) of 0.961, a root mean square error (RMSE) of 3.203 cm, and a Nash-Sutcliffe efficiency (NSE) of 0.960, outperforming the other four models. This suggests that even in areas with limited monitoring, advanced imputation methods can be useful in improving the reliability of predictions.