基于机器学习模型的数值降雨预报校正

以潮白河流域 12 个站点为研究对象，选取 12 个站点的未来 12?h 不同预见期的预报降水数据，构建基于支 持 向 量机 (support?vector?machine,?SVM) 模 型 、 随 机 森 林 (random?forest,?RF) 模 型 和 多 层 感 知 机 (multilayer perceptron，MLP) 模型的不同预见期预报降雨校正模型，模型输入为站点对应网格及其周边 8 个网格的降雨预报 数据，模型参数采用贝叶斯优化技术进行估计。利用均方根误差和确定性系数评估各模型对不同预见期预报降 水的校正效果。结果表明：未经校正的原始预报在不同预见期的预报精度均较差；各个误差校正模型在率定期与 验证期对不同预见期降雨均具有较好的校正效果；经 SVM、RF 和 MLP 模型校正后，均方根误差的平均值在率定 期分别降低了：54.2%、50.0% 和 20.8%，在验证期分别降低 42.9%、33.3% 和 14.3%；确定性系数的平均值在率定期 与验证期也均有显著提高；3 个误差校正模型中，SVM 模型表现最优，RF 模型次之。研究成果可为其他流域数值 降雨预报数据校正提供参考。

Numerical rainfall forecast correction based on machine learning model

Rainfall?is?a?direct?factor?in?the?formation?of?flood,?and?the?combination?of?accurate?rainfall?forecast?data in?the?long?forecast?period?and?hydrological?model?is?the?key?to?improve?the?accuracy?of?flood?forecast?and?increase the?forecast?period,?which?can?strive?for?a?longer?emergency?response?time?for?flood?control?and?disaster?reduction. Rainfall?forecast?data?mainly?come?from?meteorological?radar,?satellite?cloud?image?and?numerical?weather?forecast products.?Although?the?meteorological?observation?technology?and?equipment?have?made?great?progress?in?the?past few?decades,?due?to?the?chaos?of?atmospheric?system,?the?error?of?atmospheric?initial?data?and?the?error?of?model,?the rainfall ?forecast ?products ?inevitably ?have ?large ?errors ?and ?limitations, ?and ?need ?to ?be ?effectively ?corrected ?to improve?its?accuracy?and?reliability.?The?research?took?12?stations?in?Chaobai?River?basin?as?the?research?object,?the forecast?precipitation?data?of?12?stations?in?different?forecast?periods?in?the?next?12?hours?were?selected.?Rainfall error?correction?models?based?on?support?vector?machine,?random?forest?and?multilayer?perceptron?in?different forecast?periods?were?constructed.?The?model?input?is?the?rainfall?forecast?data?of?the?corresponding?grid?of?the station?and?its?surrounding?8?grids,?and?the?model?parameters?are?estimated?by?Bayesian?optimization?technology. The?root?mean?square?error?and?deterministic?coefficient?indexes?were?used?to?evaluate?the?correction?effect?of?each model?on?precipitation?forecast?in?different?forecast?periods.?The?results?showed?that?the?prediction?accuracy?of uncorrected ?original ?forecast ?was ?poor ?in ?different ?forecast ?periods. ?Each ?error ?correction ?model ?has ?a ?good correction?effect?on?rainfall?in?different?forecast?periods.?After?correction?by?support?vector?machine?model,?random forest?model?and?multilayer?perceptron?model,?the?average?root?mean?square?error?decreases?by?54.2%,?50.0%?and 20.8%,?respectively.?During?the?validation?period,?the?reduction?was?42.9%,?33.3%?and?14.3%,?respectively.?The average?certainty?coefficient?also?increased?significantly?in?both?the?rate?period?and?the?validation?period.?Among?the three?error?correction?models,?support?vector?machine?model?is?the?best,?followed?by?random?forest?model.?Based?on support?vector?machine,?random?forest?and?multi-layer?perceptron?model,?combined?with?Bayesian?optimization technology,?the?error?correction?models?of?forecast?rainfall?data?in?different?forecast?periods?were?constructed?to correct?and?analyze?the?forecast?rainfall?data?of?12?stations?in?the?Chaobai?River?basin?in?12?different?forecast periods.?The?root?mean?square?error?and?deterministic?coefficient?were?used.?The?correction?effect?is?good?and?the accuracy?of?rainfall?forecast?is?improved,?and?it?can?be?used?as?a?reference?for?the?numerical?rainfall?correction?of other?watershed?stations.