基于AdaBoost模型的大渡河流域中长期径流预报应用研究

Medium and Long Term Runoff Forecast of Daduhe River Basin Based on AdaBoost Model

以集成学习算法中的AdaBoost模型为基础,开展大渡河流域猴子岩断面未来12个月预见期(最长预见期为12个月)的中长期径流预报研究。先采用置换准确度重要性度量方法筛选各预报月份影响因子,在此基础上采用交叉验证-随机搜索方法进行模型优化,构建了各预报月份的中长期径流预报模型。通过逐月滚动建模及滚动预报的方式评估模型在不同预见期及不同月份的真实预报效果。结果表明,不同预见期(1~12月)条件下,模型预报系列与实测系列的相关系数均大于0.9,模型具有较高的预报精度。就不同预见期预报精度的综合结果来看,1~6、8~10、12月的预报精度均大于80%;7、11月的预报精度在75%左右。2018年1~12月和2020年10月至2021年9月共12个不同预见期的平均预报精度分别为85.7%、85.1%。从中长期精度预报角度而言,模型具有较好的实用精度要求,可为流域水资源精准调配和发电效益的提高提供支撑。

Based on the AdaBoost model, this paper studied on the mid-and long-term runoff forecasting for the Houziyan section of the Daduhe River Basin(the longest forecast period is 12-month). Firstly, the replacement accuracy of importance measurement method was used to select the influencing factors of each forecast month. Secondly, the cross-validation and random search method was used to optimize the model. Finally, the med-and long-term runoff forecast model for each month was constructed. The real forecast effect of the model in different forecast period was evaluated. The results show that under different forecast periods from 1-month to 12-month, the correlation coefficients between the model forecast series and the observation series are higher than 0.9, which indicates the model has high forecast accuracy. From the comprehensive results of forecast accuracy in different forecast periods, the forecast accuracy of January-June, August-October and December are all higher than 80%;The forecast accuracy of July and November is around 75%. The average forecast accuracy of 12 different forecast periods from January 2018 to December 2018 and October 2020 to September 2021 is 85.7% and 85.1%, respectively. From the perspective of medium and long-term accuracy forecasting, the model has good practical accuracy, which can provide support for the accurate allocation of water resources in the water basin and the improvement of power generation efficiency.