基于集成学习模型的非平稳月径流预报

针对变化环境下月径流序列的非平稳性日益加剧，传统径流预报模型采用普通学习算法的局限性，基于Bagging和Boosting集成学习算法，构建了随机森林(RF)、梯度提升决策树(GBDT)和轻梯度提升机(LightGBM)3种集成学习模型，融合弹性网(EN)和变分模态分解(VMD),建立VMD-EN-RF、VMD-EN-GBDT和VMD-EN-LightGBM非平稳月径流组合预报模型，并以黄河流域实测月径流为研究对象，评估预报结果的不确定性。结果表明：单一集成学习模型能够提供可靠的预报结果，适用于非平稳月径流预报；融合VMD和EN的集成学习模型预报性能较单一集成学习模型有了显著提高，纳什效率系数提升了15%～20%,均方根误差降低了30%～40%;基于Boosting集成方法构建的集成学习模型优于Bagging集成方法，其中VMD-EN-LightGBM预见期3月内的预报效果优于VMD-EN-RF和VMD-EN-GBDT,在90%置信度的区间预报覆盖率高于90%,表现出良好的性能。

Non-stationary monthly runoff forecast based on ensemble learning model

In view of the increasing non-stationary of monthly runoff series under changing environment, the limitation of ordinary learning of traditional runoff forecasting models, and the accurate runoff forecasting is facing new challenges, the research on forecasting consistent with the changing characteristics of the runoff series has been carried out. Based on Bagging and Boosting integrated learning algorithms, three integrated learning models, namely random forest (RF), gradient boosting decision tree (GBDT) and light gradient boosting machine (LightGBM), are constructed. Integrated with elastic net (EN) and variational mode decomposition (VMD), combined forecasting models of VMD-EN-RF, VMD-EN-GBDT and VMD-EN-LightGBM non-stationary monthly runoff are established. The measured monthly runoff of the Yellow River Basin is taken as the research object to evaluate the uncertainty of the forecasting results. The results show that the single ensemble learning model can provide reliable forecasting results and is suitable for complex monthly runoff forecasting. The prediction performance of the integrated learning model integrating VMD and EN is significantly improved compared with the single integrated learning model. The Nash efficiency coefficient is increased by 15%~20%, and the root mean square error is reduced by 30%~40%. The ensemble learning model built based on Boosting is better than that built based on Bagging. Among them, the prediction effect of VMD-EN-LightGBM within 3-month forecast period is better than that of VMD-EN-RF and VMD-EN-GBDT, and the coverage rate is higher than 90% in the 90% confidence interval, showing good performance.