文章摘要
基于时序分解与二维转换的集成式深度学习径流预测模型与应用
An integrated deep learning runoff prediction model based on time series decomposition and 2D transformation: methods and applications
投稿时间:2025-08-25  修订日期:2026-02-09
DOI:
中文关键词: 径流预测  集成式模型  时序分解  格拉姆角场  混沌进化优化算法  黄河流域
英文关键词: runoff prediction  ensemble model  time series decomposition  Gramian Angular Field(GAF)  Chaos-enhanced Evolutionary Optimization(CEO)  the Yellow River basin
基金项目:国家重点研发计划课题(2021YFC3201104);国家自然科学基金项目(52479001);中央高校基本科研业务费专项资金资助项目(2253200030)
作者单位邮编
徐永康 北京师范大学水科学研究院 100875
左德鹏* 北京师范大学水科学研究院 100875
韩煜娜 北京师范大学水科学研究院 
马志瑾 黄河水利委员会水文局 
刘吉峰 黄河水利委员会水文局 
徐宗学 北京师范大学水科学研究院 
摘要点击次数: 22
全文下载次数: 0
中文摘要:
      流域径流过程具有显著非平稳性与随机性,仅提供确定性结果的点预测方法,难以满足风险决策对不确定性量化的迫切需求。为系统应对这两大挑战,本研究提出一种“分解-转换-识别”集成式深度学习径流预测框架(STL-GAF-CEO-CNN-BiLSTM-ABKDE)。该框架首先采用季节趋势分解(STL)将径流序列解构为物理意义明确的子序列;随后通过格拉姆角场(GAF)将一维子序列转换为二维图像以提取其深层形态特征;然后利用卷积神经网络(CNN)与双向长短期记忆网络(BiLSTM)的混合架构进行深度识别与预测;为提升模型性能,引入混沌进化优化(CEO)算法进行超参数寻优;最后采用自适应带宽核密度估计(ABKDE)方法量化不确定性,并以黄河干流三个典型水文站月径流数据对该框架进行验证。结果表明:该集成框架在12个对比模型中表现最佳,唐乃亥、三门峡和利津站纳什效率系数(NSE)分别达0.91、0.89和0.87;在不确定性量化方面,该框架综合评价指标CWC在95%置信水平下相较于固定带宽方法降低了20.6%-29.9%,实现了区间预测可靠性与精确性的更优平衡。本研究提出的集成框架可为复杂条件下径流精准预测与不确定性量化提供新思路,对水资源风险管理具有参考价值。
英文摘要:
      Abstract: Watershed runoff processes exhibit significant non-stationarity and stochasticity. Point prediction methods, which only provide deterministic results, are inadequate for the urgent demand of uncertainty quantification in risk-based decision-making. To systematically address these two challenges, this study proposed a "Decomposition-Transformation-Identification" integrated deep learning framework for runoff prediction (STL-GAF-CEO-CNN-BiLSTM-ABKDE). This framework first employs Seasonal-Trend decomposition using LOESS (STL) to decompose the runoff series into physically meaningful sub-series. Subsequently, Gramian Angular Field (GAF) is utilized to convert the one-dimensional sub-series into two-dimensional images to extract their deep morphological features. A hybrid architecture of Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory network (BiLSTM) then performs deep recognition and prediction, with its hyperparameters optimized by a Chaos-enhanced Evolutionary Optimization (CEO) algorithm. Finally, the Adaptive Bandwidth Kernel Density Estimation (ABKDE) method is adopted to quantify prediction uncertainty. The proposed framework was validated using monthly runoff data from three typical hydrological stations on the main stream of the Yellow River. The results demonstrate that the integrated framework outperformed all 12 comparison models, achieving Nash-Sutcliffe Efficiency (NSE) coefficients of 0.91, 0.89, and 0.87 at the Tangnaihai, Sanmenxia, and Lijin stations, respectively. In terms of uncertainty quantification, the framework's comprehensive evaluation metric, the Coverage Width-based Criterion (CWC), was reduced by 20.6%-29.9% at a 95% confidence level compared to the fixed-bandwidth method, achieving a better balance between prediction reliability and accuracy in interval prediction. The integrated framework proposed in this study presents a novel approach for precise runoff prediction and uncertainty quantification under complex conditions, providing a valuable reference for water resource risk management.
  查看/发表评论  下载PDF阅读器
关闭