摘要
为应对径流预测中多尺度性和模型复杂性对实时预测的影响,提出一种新型径流预测模型CEEMDANPyraformer-LSS,以确保模型的鲁棒性。利用自适应噪声完备集合经验模态分解(complete ensemble empirical mode decomposition with adaptive noise,CEEMDAN)技术将复杂径流序列分解为多个本征模态函数,有效去除噪声并提高数据平稳性。利用金字塔注意力模型(Pyraformer)的多头自注意力机制对这些分解后的数据进行高效特征提取和预测。为增强模型在处理噪声和突发事件时的鲁棒性和适应性,结合局部随机敏感性(localized stochastic sensitivity,LSS)函数,动态调整对最新数据点和异常值的敏感度。针对永定河4个关键断面开展应用研究,结果表明:CEEMDAN-Pyraformer-LSS模型在径流预测中精度达到95%。与LSTM、BP模型相比,基于Pyraformer的预测模型在水资源管理和防洪预警等实际应用场景实现了高效性和鲁棒性。
Rainfall-runoff prediction processes has become crucial for water resources planning and management with the increasing impact of climate change and human activities,especially for the real-time forecasting of reservoir inflows.Real-time and robust runoff predictions are essential to ensure effective water resource allocation and flood prevention strategies.Yongding River,one of the seven major rivers in the Haihe River basin,serves as an important case study for this research.Analyzing the historical runoff data and establishing prediction models can be used to effectively forecast the future runoff trends,which are vital for flood prevention,water resource management,and ecological protection.The complex terrain and variable hydrological conditions of the downstream Yongding River necessitates a model with high adaptability and flexibility.In practical applications,due to the demands of water resource management and flood warning,quick decision-making is required,which necessitates not only maintaining the accuracy of the model,but also improving its real-time performance and robustness.To address the impact of multi-scale characteristics and model complexity on real-time runoff prediction,this study proposes a novel runoff prediction model-CEEMDAN-Pyraformer-LSS.First,the complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)decomposition technique is employed to decompose complex runoff sequences into multiple intrinsic mode functions,effectively removing noise and enhancing data stability.This preprocessing step ensures that subsequent analysis focuses on the significant components of the runoff data.After decomposition,the pyramid attention model(Pyraformer)with its multi-head self-attention mechanism is used for efficient feature extraction and prediction of the decomposed data.The Pyraformer network is designed to capture long-range dependencies and multi-scale temporal relationships within the data,which are crucial for real-time runoff prediction.The model's architecture allows for efficient processing of large datasets and extraction of meaningful patterns essential for runoff prediction.A local stochastic sensitivity(LSS)function is integrated to further enhance the model's robustness and adaptability in handling noise and sudden events.This function dynamically adjusts the sensitivity to the latest data points and anomalies,improving the model's performance in noisy and unpredictable environments.The LSS function enables the model to effectively respond to sudden changes in the data,such as unexpected rainfall or emergency reservoir measures.The CEEMDAN-Pyraformer-LSS model achieves accurate and efficient real-time predictions through effective feature extraction and noise processing,with a prediction accuracy of 95%.The Pyraformer network demonstrates excellent predictive performance across different sliding window lengths,attention layers,and attention heads,achieving optimal results when set to 4,4,and 30,respectively.Its pyramid attention mechanism significantly reduces model complexity through multi-level feature extraction and information integration.The incorporation of the CEEMDAN decomposition model increased the Nash-sutcliffe efficiency(ENS)of the prediction results by 6.39% and reduced the root mean square error(ERMS)by 35%,showcasing its exceptional ability in noise reduction and enhancing data stability.The introduction of the LSS function further improved the model's sensitivity and robustness during peaks and abrupt changes.Ultimately,the model achieved an error rate of 11.59% and a correlation coefficient of 99.74%.The above analysis fully verifies the superiority of the robust attention model based on the Pyraformer network and its effectiveness in addressing the real-time runoff prediction challenges of the Yongding River.The CEEMDAN-Pyraformer-LSS model not only ensures high predictive accuracy,but also demonstrates exceptional adaptability and stability in dynamic and noisy environments.The integration of decomposition techniques,attention mechanisms,and adaptive sensitivity functions within the model effectively tackles the issues posed by the multiscale nature of runoff data and model complexity,significantly enhancing the robustness of real-time predictions.This highlights the model's crucial role in effective water resource management and flood prevention.
作者
孙祥瑜
王超
杨一旸
张利娜
蔡思宇
康龙熙
王浩
SUN Xiangyu;WANG Chao;YANG Yiyang;ZHANG Lina;CAI Siyu;KANG Longxi;WANG Hao(Jiangsu University,Zhenjiang 212013,China;China Institute of Water Resources and Hydropower Research,Beijing 100038,China;Xiamen University,Xiamen 361005,China;Northeastern University,Shenyang 110057,China)
出处
《南水北调与水利科技(中英文)》
北大核心
2025年第2期363-374,共12页
South-to-North Water Transfers and Water Science & Technology
基金
国家重点研发计划项目(2022YFC3204603)
水工程智能调度控制技术装备与仿真测试平台研发项目(SKS-2022117)。
关键词
降雨径流
实时预测
机器学习
金字塔注意力模型
局部随机敏感性
rainfall-runoff
dynamic real-time prediction
machine learning
pyramid attention model
localized stochastic sensitivity
