Scholarlink

<p>Groundwater is crucial, especially in arid regions. Predicting groundwater levels (GWLs) is influenced by factors like rainfall and soil permeability. This study introduces a novel approach using a multi-branch spatio-temporal graph convolutional neural network (STGCN-M) to predict GWLs, modeled as an undirected graph problem. By constructing graphs from the locations of observation wells and applying sliding window techniques, this model captures dynamic groundwater fluctuations effectively. It successfully predicted GWLs for wells in the Yinchuan alluvial plain in the northern Ningxia Hui Autonomous Region of China Plain from 1990 to 2018 and the hourly GWLs of 52 observation wells from July 2021 to September 2021, showing promising results. Experimental results indicate that in the different experimental groups of the hourly dataset of the Yinchuan Plain (training set ratio 60%/70%/80%), the STGCN (RMSE(root mean square error) = 0.01561/0.01136/0.01100) and STGCN-M (RMSE = 0.00942/0.00860/0.00790) is significantly better than LSTM (RMSE = 0.93444/0.64135/0.50616) and GRU (RMSE = 0.58712/0.59100/0.45158). Additionally, the RMSE of the multi-branch STGCN-M is 41.58%, 24.30%, and 28.18% higher than that of the single-branch STGCN in different experimental groups, respectively. The proportion of observation wells in which STGCN-M is better than STGCN is 96.2%, 73.1%, and 80.8%, respectively. Moreover, in the different groups of experiments on the monthly datasets, the overall prediction performance of STGCN-M was the best (NSE(Nash–Sutcliffe model efficiency coefficient) = 0.66.</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 the accuracy of groundwater level prediction at different scales using spatio-temporal graph convolutional model

  • 期刊论文

  • 发表日期：

    2025年2月11日

查看全文
  • Long Chen,
  • Dezheng Zhang,
  • Jianwei Xu,
  • Zikun Zhou,
  • Jianing Jin,
  • Jing Luan,
  • Aziguli Wulamu

摘要

Groundwater is crucial, especially in arid regions. Predicting groundwater levels (GWLs) is influenced by factors like rainfall and soil permeability. This study introduces a novel approach using a multi-branch spatio-temporal graph convolutional neural network (STGCN-M) to predict GWLs, modeled as an undirected graph problem. By constructing graphs from the locations of observation wells and applying sliding window techniques, this model captures dynamic groundwater fluctuations effectively. It successfully predicted GWLs for wells in the Yinchuan alluvial plain in the northern Ningxia Hui Autonomous Region of China Plain from 1990 to 2018 and the hourly GWLs of 52 observation wells from July 2021 to September 2021, showing promising results. Experimental results indicate that in the different experimental groups of the hourly dataset of the Yinchuan Plain (training set ratio 60%/70%/80%), the STGCN (RMSE(root mean square error) = 0.01561/0.01136/0.01100) and STGCN-M (RMSE = 0.00942/0.00860/0.00790) is significantly better than LSTM (RMSE = 0.93444/0.64135/0.50616) and GRU (RMSE = 0.58712/0.59100/0.45158). Additionally, the RMSE of the multi-branch STGCN-M is 41.58%, 24.30%, and 28.18% higher than that of the single-branch STGCN in different experimental groups, respectively. The proportion of observation wells in which STGCN-M is better than STGCN is 96.2%, 73.1%, and 80.8%, respectively. Moreover, in the different groups of experiments on the monthly datasets, the overall prediction performance of STGCN-M was the best (NSE(Nash–Sutcliffe model efficiency coefficient) = 0.66.