基于机器学习模型的淮河流域中长期径流预报研究

准确可靠的中长期径流预报是支撑水资源科学调配、提高水资源利用效率的关键。本研究采用AdaBoost模型(AdB)、随机森林模型(RF)和支持向量机模型(SVM)进行淮河流域王家坝和蚌埠站当年11月至次年10月共12个月的中长期径流预报研究。采用置换准确度重要性度量法从130项气象-气候因子及前期降雨/流量构建的1 562个因子变量中筛选出影响各月径流的关键因子,构建了基于AdB、RF和SVM模型的各月径流预报模型,模型参数采用随机搜索技术并结合交叉验证方式确定。采用变幅误差合格率和等级(五级)预报合格率指标对模型的预报精度进行了评估。变幅误差合格率指标表明,王家坝12个月的平均合格率分别为99.8%(AdB)、96.6%(RF)和95.9%(SVM),蚌埠站分别为100%(AdB)、94.8%(RF)和93.8%(SVM);等级预报合格率指标表明,王家坝12个月的平均合格率分别为79.0%(AdB)、76.4%(RF)和79.9%(SVM),蚌埠站分别为81.0%(AdB)、75.6%(RF)和76.6%(SVM)。模型均具有较好的预报效果,但RF和SVM模型对于高流量值的预报存在偏低现象,AdB模型整体上优于RF和SVM模型。

Medium to long term runoff forecast for the Huai River Basin based on machine learning algorithm

Accurate and reliable medium to long term runoff prediction is key to supporting scientific allocation of water resources and improving water resources utilization efficiency. In this study, the AdaBoost (AdB), Random forest (RF) and Support vector machine (SVM) algorithms were used to make medium to long term runoff predictions at the Wangjiaba and Bengbu stations, Huai River Basin, from November to October of the following year. The permutation accuracy importance measure was used to select the key factors affecting monthly runoff from 1562 factor variables constructed from 130 meteorological-climate factors and previous precipitation/runoff fluxes, and the monthly runoff forecasts based on the aforementioned three machine learning algorithms were made for each month, using model parameters determined by way of random search combined with cross-validation. The performances of the three algorithms were evaluated using the variable amplitude error qualification rate indicator and the grade (five level) forecast qualification rate indicator. The former indicator shows that the 12-month average qualification rates of prediction for the three algothrims at the Wangjiaba station were, respectively, 99.8% (AdB), 96.6% (RF) and 95.5% (SVM) and at the Bengbu station 100% (AdB), 94.8% (RF) and 93.8% (SVM); the results using the latter indicator at the Wangjiaba station were, respectively, 79.0% (AdB), 76.4% (RF) and 79.9% (SVM), and at the Bengbu Station 81.0% (AdB), 75.6% (RF) and 76.6% (SVM). The three machine learning algorithms all performed well, but RF and SVM had low prediction rates on high discharge values, and AdB performed better than RF and SVM overall.