Scholarlink

<p>Previous studies have employed the water level information at the individual flood gauging stations to perform predictions. However, it is widely recognized that the water levels at downstream stations are correlated with those at upstream stations. Therefore, the innovation and contribution of this study is the proposal of a new spatiotemporal watershed-scale machine learning (ML) model for hourly and daily water level prediction in the tidal Beigang River, Taiwan, considering the correlation between flood gauging stations. The proposed new model is based on a novel integration of a primary learner of a spatial–temporal graph convolutional network, an enhancing learner, and residual error corrections. By extracting spatiotemporal features from a hydrological network, the measured ten-year hydrological data (2012–2022) were employed. Three multioutput (MO)-based ML models were presented and utilized as benchmarks, including extra tree regression (MO-ETR), light gradient boosting machine regression (MO-LGBMR), and support vector regression (MO-SVR). The SHapley Additive exPlanations (SHAP) method was used to investigate the global contribution of inputs to the water level predictions. Two existing encoder-decoder (ED)-based deep learning (DL) models, namely gated recurrent units (ED-GRU) and long short‐term memory (ED-LSTM), were employed to examine the improved prediction performance of proposed model. Compared to the three MO-based and two ED-based DL models, the proposed model is more feasible, deterministic, and accurate. Regarding the hourly water level prediction performance, the proposed model exhibited a notable improvement in the total-averaged Nash–Sutcliffe efficiency (NSE) performance when compared to the MO-ETR, MO-LGBMR, MO-SVR, ED-LSTM, and ED-GRU models, with a percentage enhancement of 49.2%, 31.4%, 29.2%, 33.3%, and 32.2%, respectively. This study further provided a test study with insights from an additional Cho-Shui River basin, with the objective of better investigating the applicability of the proposed model. Satisfactory and reliable results suggest that the proposed model significantly contributes to and serves as a reference for forecasting the evolution of river floods.</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

A spatiotemporal watershed-scale machine-learning model for hourly and daily flood-water level prediction: the case of the tidal Beigang River, Taiwan

  • 期刊论文

  • 发表日期：

    2025年3月3日

查看全文
  • Wen-Dar Guo,
  • Wei-Bo Chen,
  • Chih-Hsin Chang

摘要

Previous studies have employed the water level information at the individual flood gauging stations to perform predictions. However, it is widely recognized that the water levels at downstream stations are correlated with those at upstream stations. Therefore, the innovation and contribution of this study is the proposal of a new spatiotemporal watershed-scale machine learning (ML) model for hourly and daily water level prediction in the tidal Beigang River, Taiwan, considering the correlation between flood gauging stations. The proposed new model is based on a novel integration of a primary learner of a spatial–temporal graph convolutional network, an enhancing learner, and residual error corrections. By extracting spatiotemporal features from a hydrological network, the measured ten-year hydrological data (2012–2022) were employed. Three multioutput (MO)-based ML models were presented and utilized as benchmarks, including extra tree regression (MO-ETR), light gradient boosting machine regression (MO-LGBMR), and support vector regression (MO-SVR). The SHapley Additive exPlanations (SHAP) method was used to investigate the global contribution of inputs to the water level predictions. Two existing encoder-decoder (ED)-based deep learning (DL) models, namely gated recurrent units (ED-GRU) and long short‐term memory (ED-LSTM), were employed to examine the improved prediction performance of proposed model. Compared to the three MO-based and two ED-based DL models, the proposed model is more feasible, deterministic, and accurate. Regarding the hourly water level prediction performance, the proposed model exhibited a notable improvement in the total-averaged Nash–Sutcliffe efficiency (NSE) performance when compared to the MO-ETR, MO-LGBMR, MO-SVR, ED-LSTM, and ED-GRU models, with a percentage enhancement of 49.2%, 31.4%, 29.2%, 33.3%, and 32.2%, respectively. This study further provided a test study with insights from an additional Cho-Shui River basin, with the objective of better investigating the applicability of the proposed model. Satisfactory and reliable results suggest that the proposed model significantly contributes to and serves as a reference for forecasting the evolution of river floods.