Yao Huaipeng, Huang Yuling, Lv Pingyu, Luo Huihuang
School of River and Ocean Engineering, Chongqing Jiao Tong University, Chongqing, 400074, China.
China Institute of Water Resources and Hydropower Research, Beijing, 100038, China.
Sci Rep. 2024 Aug 21;14(1):19442. doi: 10.1038/s41598-024-70262-4.
Accurate and rapid prediction of water quality is crucial for the protection of aquatic ecosystems. This study aims to enhance the prediction of total phosphorus (TP) concentrations in the middle reaches of the Yangtze River by integrating advanced modeling techniques. Using operational and discharge data from the Three Gorges Reservoir (TGR), along with water quality parameters from downstream sections, we used Grey Relational Analysis (GRA) to rank the factors contributing to TP concentrations. The analysis identified turbidity, permanganate index (CODMn), total nitrogen (TN), water temperature, chlorophyll a, upstream water level variation, and discharge from the Three Gorges Dam (TGD) as the top contributors. Subsequently, a coupled neural network model was established, incorporating these key contributors, to predict TP concentrations under the dynamic water level control during flood periods in the TGR. The proposed GRA-CEEMDAN-CN1D-LSTM-DBO model was compared with conventional models, including BP, LSTM, and GRU. The results indicated that the GRA-CEEMDAN-CN1D-LSTM-DBO model significantly outperformed the others, achieving a correlation coefficient (R) of 0.784 and a root mean square error (RMSE) of 0.004, compared to 0.58 (R) and 0.007 (RMSE) for the LSTM model, 0.576 (R) and 0.007 (RMSE) for the BP model, and 0.623 (R) and 0.006 (RMSE) for the GRU model. The model's accuracy and applicability further validated in two sections: YC (Yunchi) in Yichang City and LK (Liukou) in Jingzhou City, where it performed satisfactorily in predicting TP in YC (R = 0.776, RMSE = 0.007) and LK (R = 0.718, RMSE = 0.007). Additionally, deep learning analysis revealed that as the distance away from dam increased, prediction accuracy gradually decreased, indicating a reduced impact of TGR operations on downstream TP concentrations. In conclusion, the GRA-CEEMDAN-CN1D-LSTM-DBO model demonstrates superior performance in predicting TP concentration in the middle reaches of the Yangtze River, offering valuable insights for dynamic water level control during flood seasons and contributing of smart to the advancement of water management in the Yangtze River.
准确快速地预测水质对于保护水生生态系统至关重要。本研究旨在通过整合先进的建模技术，提高长江中游总磷（TP）浓度的预测能力。利用三峡水库（TGR）的运行和流量数据，以及下游断面的水质参数，我们采用灰色关联分析（GRA）对影响TP浓度的因素进行排序。分析确定浊度、高锰酸盐指数（CODMn）、总氮（TN）、水温、叶绿素a、上游水位变化和三峡大坝（TGD）的流量是主要贡献因素。随后，建立了一个耦合神经网络模型，纳入这些关键贡献因素，以预测三峡水库汛期动态水位控制下的TP浓度。将所提出的GRA-CEEMDAN-CN1D-LSTM-DBO模型与传统模型（包括BP、LSTM和GRU）进行了比较。结果表明，GRA-CEEMDAN-CN1D-LSTM-DBO模型明显优于其他模型，相关系数（R）为0.784，均方根误差（RMSE）为0.004，而LSTM模型的R为0.58，RMSE为0.007；BP模型的R为0.576，RMSE为0.007；GRU模型的R为0.623，RMSE为0.006。该模型的准确性和适用性在宜昌市的YC（云池）和荆州市的LK（柳口）两个断面进一步得到验证，在预测YC断面的TP时表现令人满意（R = 0.776，RMSE = 0.007），在预测LK断面的TP时（R = 0.718，RMSE = 0.007）也是如此。此外，深度学习分析表明，随着离大坝距离的增加，预测精度逐渐降低，这表明三峡水库运行对下游TP浓度的影响减小。总之，GRA-CEEMDAN-CN1D-LSTM-DBO模型在预测长江中游TP浓度方面表现出卓越性能，为汛期动态水位控制提供了有价值的见解，并为长江水资源管理的智能化发展做出了贡献。
