Scholarlink

<p>Well water level is an important precursor observation, which is expected to be used to extract information on subsurface stress and media changes. Real-time prediction of well water level can help prevent geological disasters, but there are few related experimental studies. This study aims to explore a short-term prediction model of well water level that is more pervasive than the GRU model, explore new methods to enhance the model’s capability, and provide scientific references for the application of deep learning models in the field of well water level prediction. Taking the measured data of the Three Gorges well network from 2012 to 2014 as an example, the performance of the GRU and its variant models on the RMSE, MAE and R² evaluation criteria are compared, and the results show that only the BiGRU-Attention model shows excellent performance at all well points, with better pervasiveness and stability; performing a single-step prediction and adding a 1% standard deviation noise to the training set can improve the robustness and generalisation of the model.</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

Micro-dynamics prediction of well water level based on GRU and attention mechanism

  • 期刊论文

  • 发表日期：

    2025年9月6日

查看全文
  • Xiaoyu Fang,
  • Lili Zhang,
  • Haoran Li,
  • Yaowen Zhang,
  • Yunsheng Yao

摘要

Well water level is an important precursor observation, which is expected to be used to extract information on subsurface stress and media changes. Real-time prediction of well water level can help prevent geological disasters, but there are few related experimental studies. This study aims to explore a short-term prediction model of well water level that is more pervasive than the GRU model, explore new methods to enhance the model’s capability, and provide scientific references for the application of deep learning models in the field of well water level prediction. Taking the measured data of the Three Gorges well network from 2012 to 2014 as an example, the performance of the GRU and its variant models on the RMSE, MAE and R² evaluation criteria are compared, and the results show that only the BiGRU-Attention model shows excellent performance at all well points, with better pervasiveness and stability; performing a single-step prediction and adding a 1% standard deviation noise to the training set can improve the robustness and generalisation of the model.