机器学习模型在H-ADCP在线测流系统中的应用

Application of machine learning models in streamflow online monitoring system using H-ADCP

针对常规的指标流速法精度不足的缺点,分析探究了利用机器学习模型由H-ADCP网格单元流速推求测流断面平均流速的可行性与适用性。选择引丹灌渠清泉沟隧洞出口下游清泉沟站2018~2019年共136测次人工比测流量资料及同时期H-ADCP网格单元流速资料,分别构建了3种机器学习模型(支持向量回归、BP神经网络和极限学习机),并基于8种H-ADCP网格单元分配方案对断面平均流速进行了模拟。以指标流速法为参照基准,对比分析了3种机器学习模型的拟合效果。结果表明:相比常规的指标流速法,机器学习模型能更精确地拟合断面平均流速值。此外,H-ADCP有效网格单元数对指标流速法的断面平均流速拟合效果影响较大,对各机器学习方法的拟合性能影响不显著。研究成果可为机器学习模型与传统水文测验方法的深度融合提供新的研究思路。

The index velocity method has been widely used to estimate a river’s mean water velocity by H-ADCP,but its insufficient accuracy remains an issue.Thus,this study aims to explore the potential to provide more accurate estimation of river’s mean water velocity by machine learning model.The investigation was carried out in Qingquangou Station,where the observed streamflow data series(136 measurements)and corresponding H-ADCP measurements were obtained for the period 2018~2019.Three machine learning models(support vector regression,BP neural network and extreme learning machine)were established and trained under 8 training schemes where different numbers of H-ADCP cells were assigned.Their simulation performances of mean flow velocity were then assessed with the index velocity method as benchmark.The results showed that the machine learning models could achieve a better estimation of mean flow velocity when compared to the traditional index velocity method.Besides,the simulation performance of the index velocity method highly depended on the number of effective cells of H-ADCP,while the machine learning models showed a weaker response to the possible failure of certain cell of H-ADCP.The paper might provide new insights for incorporating machine learning technique into the practice of traditional hydrological measurements.