Scholarlink

<p>Accurate chlorophyll-a (Chl-a) forecasting is essential for early warning systems of algal blooms in coastal waters. This task remains challenging due to the pronounced nonlinearities inherent in phytoplankton dynamics and persistent observational noise. This study proposes CEEMDAN-SF-GAGRU, a hybrid approach that integrates complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), a smoothing filter (SF), and a genetic algorithm-optimized gated recurrent unit (GAGRU). The proposed method was evaluated against eight benchmark models using observational data collected at two coastal monitoring sites. The benchmark set included four individual models: backpropagation neural network (BPNN), generalized regression neural network (GRNN), least squares support vector machine (LSSVM), and GAGRU. Additionally, four CEEMDAN-based hybrid models were tested: CEEMDAN-BPNN, CEEMDAN-GRNN, CEEMDAN-LSSVM, and CEEMDAN-GAGRU. Results demonstrate that CEEMDAN-SF preprocessing effectively reduces noise and irregular fluctuations, resulting in substantial improvements in forecasting accuracy. For 1&#xa0;day step-ahead forecasting, CEEMDAN-SF-GAGRU demonstrated superior performance. The method achieved coefficients of determination of 0.979 and 0.921 at the two monitoring stations. Furthermore, it reduced the root-mean-square error by more than 28% compared to the best-performing benchmark model. The proposed method also demonstrated superior capability in tracking Chl-a trends and peak values, showing enhanced robustness under conditions of extreme concentration variations. This study presents a robust and reliable approach for forecasting Chl-a concentrations in coastal waters, offering valuable insights for algal bloom monitoring and management. Highlights. The integration of CEEMDAN and SF methods effectively reduces noise interference and non-stationary components during model training. The CEEMDAN-SF-GAGRU method demonstrates stronger performance in capturing trends and peak values of chl-a concentrations. The new hybrid method demonstrates stronger adaptability to different forecasting steps ahead compared to other methods.</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

Forecasting of Chlorophyll-a Concentrations in Coastal Waters Through Gated Recurrent Unit Neural Network Based on Time–frequency Analysis and Smoothing Filtering

  • 期刊论文

  • 发表日期：

    2026年3月9日

查看全文
  • Qin Ye,
  • Zhongsheng Yi,
  • Yueji Liang,
  • Xiaoting Huang,
  • Jing Li

摘要

Accurate chlorophyll-a (Chl-a) forecasting is essential for early warning systems of algal blooms in coastal waters. This task remains challenging due to the pronounced nonlinearities inherent in phytoplankton dynamics and persistent observational noise. This study proposes CEEMDAN-SF-GAGRU, a hybrid approach that integrates complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), a smoothing filter (SF), and a genetic algorithm-optimized gated recurrent unit (GAGRU). The proposed method was evaluated against eight benchmark models using observational data collected at two coastal monitoring sites. The benchmark set included four individual models: backpropagation neural network (BPNN), generalized regression neural network (GRNN), least squares support vector machine (LSSVM), and GAGRU. Additionally, four CEEMDAN-based hybrid models were tested: CEEMDAN-BPNN, CEEMDAN-GRNN, CEEMDAN-LSSVM, and CEEMDAN-GAGRU. Results demonstrate that CEEMDAN-SF preprocessing effectively reduces noise and irregular fluctuations, resulting in substantial improvements in forecasting accuracy. For 1 day step-ahead forecasting, CEEMDAN-SF-GAGRU demonstrated superior performance. The method achieved coefficients of determination of 0.979 and 0.921 at the two monitoring stations. Furthermore, it reduced the root-mean-square error by more than 28% compared to the best-performing benchmark model. The proposed method also demonstrated superior capability in tracking Chl-a trends and peak values, showing enhanced robustness under conditions of extreme concentration variations. This study presents a robust and reliable approach for forecasting Chl-a concentrations in coastal waters, offering valuable insights for algal bloom monitoring and management. Highlights. The integration of CEEMDAN and SF methods effectively reduces noise interference and non-stationary components during model training. The CEEMDAN-SF-GAGRU method demonstrates stronger performance in capturing trends and peak values of chl-a concentrations. The new hybrid method demonstrates stronger adaptability to different forecasting steps ahead compared to other methods.