Scholarlink

<p>Accurate nitrate estimation in surface water is essential for ecological and human health. To predict nitrate levels (NO<sub>3</sub>; mg/L) in the Willamette River at Portland, Oregon, USA, three stand-alone deep learning models—long short-term memory (LSTM), gated recurrent unit (GRU), and convolutional neural network (CNN)—as well as their new hybrid models, CNN-LSTM and CNN-GRU, were developed based on nine water quality parameters: river discharge (Q), dissolved oxygen (DO), turbidity (TU), mean water velocity (V), specific conductance (SC), water temperature (T), pH, gage height (GH), and chlorophyll (Chl) for the period 2021–2022 (two years). Five different temporal scenarios were analyzed, corresponding to time intervals of 1, 3, 6, 12, and 24 h. The performance of models was assessed using the correlation coefficient (R), Nash–Sutcliffe efficiency (NSE), and root mean squared error (RMSE). DO showed the strongest correlation with NO<sub>3</sub>, and parameters were added based on their correlation strength. The N(V) model, which incorporates Q, DO, TU, V, SC, and T, yielded the highest prediction accuracy. The best results were obtained in the 1-h scenario. Results indicated that hybrid models CNN-LSTM (R = 0.892, NSE = 0.795, RMSE = 0.079) and CNN-GRU (R = 0.876, NSE = 0.761, RMSE = 0.086) outperformed stand-alone models. CNN showed superior performance among the individual models. In the sensitivity analysis, removing the turbidity parameter resulted in a 26.41% decrease in the NSE value and a 43.03% increase in the RMSE, demonstrating that turbidity is the most significant parameter for predicting NO<sub>3</sub> levels.</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

Innovative approaches to surface water quality management: advancing nitrate (NO3) forecasting with hybrid CNN-LSTM and CNN-GRU techniques

  • 期刊论文

  • 发表日期：

    2025年1月20日

查看全文
  • Sina Davoudi,
  • Kiyoumars Roushangar

摘要

Accurate nitrate estimation in surface water is essential for ecological and human health. To predict nitrate levels (NO3; mg/L) in the Willamette River at Portland, Oregon, USA, three stand-alone deep learning models—long short-term memory (LSTM), gated recurrent unit (GRU), and convolutional neural network (CNN)—as well as their new hybrid models, CNN-LSTM and CNN-GRU, were developed based on nine water quality parameters: river discharge (Q), dissolved oxygen (DO), turbidity (TU), mean water velocity (V), specific conductance (SC), water temperature (T), pH, gage height (GH), and chlorophyll (Chl) for the period 2021–2022 (two years). Five different temporal scenarios were analyzed, corresponding to time intervals of 1, 3, 6, 12, and 24 h. The performance of models was assessed using the correlation coefficient (R), Nash–Sutcliffe efficiency (NSE), and root mean squared error (RMSE). DO showed the strongest correlation with NO3, and parameters were added based on their correlation strength. The N(V) model, which incorporates Q, DO, TU, V, SC, and T, yielded the highest prediction accuracy. The best results were obtained in the 1-h scenario. Results indicated that hybrid models CNN-LSTM (R = 0.892, NSE = 0.795, RMSE = 0.079) and CNN-GRU (R = 0.876, NSE = 0.761, RMSE = 0.086) outperformed stand-alone models. CNN showed superior performance among the individual models. In the sensitivity analysis, removing the turbidity parameter resulted in a 26.41% decrease in the NSE value and a 43.03% increase in the RMSE, demonstrating that turbidity is the most significant parameter for predicting NO3 levels.