The Study Of The Mid-long Term Hydrological Forecasting Based On Ensemble Learning

Our country suffered flood disaster frequently in recent years,which caused great loss of economic and social development.The mid-long term hydrological forecasting which has longer forecast time is important basic hydraulic engineering and non-engineering measure for disaster prevention and reduction.We can take measures as early as possible to minimize losses and maximize gains in flood prevention and other water resource problems with high-accuracy mid-long term hydrological forecasting.However traditional methods hardly meet the accuracy requirements according to the complexity of hydrological system.So it becomes more and more important to introduce new methods and improve accuracy of mid-long term hydrological forecasting.The new method,namely ensemble method,which is widely applied in machine learning,is introduced to the study of mid-long term hydrological forecasting,by summarizing mid-long term hydrological forecasting models at home and abroad and analyzing historical runoff in Hutiaoaxia dam area.150 forecast factors of atmospheric circulation index and historical runoff are selected on mutual information.The respective Gradient Boosting Regression Tree(GBRT),Random Forest(RF)algorithms and combination methods combine base learner(Regression Tree)to gain generalization of algorithms and reduce risk of over-fitting effectively.In the case study,the model,which is trained on data from 1959 to 1992,predicts data from 1993 to 2000 after parameter tuning,and compares its results with single learner SVM.Then 10 most important features are selected by the property that the base learner of GBRT and RF picks out important nodes.The results suggest that the two different ensemble methods predict better than SVM,and achieve high accuracy in non-flood season.Furthermore,two methods have advantages and disadvantages in different measurement.RF predicts more excellent results than GBRT,while has higher overall average error.Combination method which combines GBRT,RF and SVM with weighted average predicts more excellent results,while shows no superiority on other indicators.Compared with Linear Regression after feature selecting,different ensemble methods have lower relative error and nearly the same mean square error.Meanwhile the results become worse after selecting.The results also inspire the study of forecast factors.