基于深度学习的黄河下游水流阻力计算研究

Flow resistance calculation in lower Yellow River based on deep learning

摘要

【目的】冲积河流阻力的准确计算在治河防洪工程中具有重要意义。传统阻力计算公式与现有机器学习方法仍存在诸多不足。为提升阻力模型的性能与泛化能力,建立一套基于深度学习的阻力计算方法,【方法】选择弗劳德数、体积含沙量、宽深比、径深比、年径流量和年输沙量等水文特征作为模型输入,构建一种基于深度森林的水流阻力计算模型。利用黄河下游水文站的实测数据进行模型训练与测试,并从时空泛化能力、迁移学习表现等方面综合评估模型。【结果】模型在测试集上的纳什效率(NSE)为0.785,平均绝对误差(MAE)为0.002,均方根误差(RMSE)为0.003,平均绝对百分比误差(MAPE)为14.618%。加入时空平均特征泛化模型后,模型的NSE从0.681 4提升至0.742 7, MAE从0.002 3降至0.002 1, RMSE从0.003 2降至0.002 8, MAPE从14.978%降至13.689%。在将模型迁移至全新的水沙条件下时, NSE最大降幅达65.35%, MAE、 RMSE和MAPE的最大升幅分别为100%、150%和123.98%。【结论】深度森林模型相比于传统的阻力公式和机器学习模型,在冲积河流一般条件下的阻力计算方面,能展现出更为优越的预测精度。通过引入大尺度时空平均特征,模型在不同水文站与水文时期的计算精度有效提升,泛化能力明显增强。然而,在应对特殊水沙情势时,模型仍存在性能波动,有时计算精度还超不过具有物理背景的糙率公式,亟需结合物理机制以增强其迁移学习能力,且在解决复杂环境下的黄河动床阻力计算问题时,应该注重同良好的传统公式相互印证这一环节。

Abstract

[Objective] Accurate calculation of flow resistance in alluvial rivers is of great significance for river regulation and flood control engineering. Conventional resistance formulas and existing machine learning method still have multiple limitations.To improve the performance and generalization ability of resistance models, a flow resistance estimation method based on deep learning is proposed. [Methods] Hydrological features, including Froude number, volumetric sediment concentration, width-todepth ratio, diameter-to-depth ratio, annual runoff, and annual sediment load, were selected as model inputs, and a flow resistance calculation model based on deep forest was established. The model was trained and tested using measured data from hydrological stations in the lower Yellow River, and comprehensively evaluated in terms of spatiotemporal generalization ability and transfer learning performance. [Results] The model achieved a Nash-Sutcliffe efficiency(NSE) of 0. 785, a mean absolute error(MAE) of 0. 002, a root mean square error(RMSE) of 0. 003, and a mean absolute percentage error(MAPE) of 14. 618% on the test dataset. After the incorporation of spatiotemporal average features, the NSE of the model increased from 0. 681 4 to 0. 742 7, the MAE decreased from 0. 002 3 to 0. 002 1, the RMSE dropped from 0. 003 2 to 0. 002 8, and the MAPE reduced from 14. 978% to 13. 689%. When the model was transferred to completely new water-sediment conditions, the maximum decline in NSE reached 65. 35%, and the maximum increases in MAE, RMSE, and MAPE were 100%, 150%, and 123. 98%, respectively. [Conclusion] Compared with traditional resistance formulas and machine learning method, the deep forest model demonstrates superior accuracy in predicting flow resistance under general conditions in alluvial rivers. By introducing large-scale spatiotemporal average features, the model's calculation accuracy across different hydrological stations and hydrological periods is effectively improved, and its generalization ability is significantly enhanced. However, under special water-sediment conditions, the model still shows performance fluctuations. In certain cases, its calculation accuracy is even lower than that of physically based roughness formulas. Therefore, it is urgent to incorporate physical mechanisms to enhance its transfer learning capability. When addressing the calculation of movable bed resistance in the Yellow River under complex environments, emphasis should be placed on mutual verification with reliable traditional formulas.

关键词

黄河下游 / 水流阻力 / 深度森林 / 迁移学习 / 弗劳德数 / 径流 / 输沙量 / 机器学习模型

Key words

lower Yellow River / flow resistance / deep forest / transfer learning / Froude number / runoff / sediment transport volume / machine learning models

基金资助

国家重点研发计划项目(2023YFC3208603)