Study On Mid-long Term Runoff Forecasting Method Based On Multi-model Forecasting Information Fusion

Water resource is a kind of precious resource.Where there is water,there is life.And the life and production of modern society are inseparable from water.People must deal with the increasingly serious problem of water resources.It is of great significance to study the evolution trend of runoff and make scientific prediction for the development and utilization of water resources,the safe and efficient operation of water conservancy facilities and disaster prevention and mitigation.In this paper,the present situation of mid-long term runoff prediction and the principles of several process-driven models and data-driven models are described in detail.Then the methods and theories of runoff characteristic analysis are studied.The basic composition,trend,jump and period of runoff time series are analyzed,which lays a foundation for mid-long term runoff forecast.The main research and innovative results of the paper are as follows:(1)This paper starts with the important aspect of establishing mid-long term runoff forecast,that is,the selection of forecast factors,and selects three methods of GRA,RF and MI to select three groups of different forecast factors.Three groups of prediction factors were input into five basic models,namely,deep neural network model,Elman neural network model,BP neural network model,multiple regression model and support vector machine model.The results show that the factors screened by GRA and MI are better than those screened by RF.(2)Then,based on the idea of ensemble learning and multi model combination,a total of 16 ensemble learning models are constructed with Stacking algorithm as the framework,deep neural network model,Elman neural network model and BP neural network model as the base model,and Ada Boost,GBDT,XGboost and RF algorithm as the meta model.At the same time,the traditional combination model based on multiple regression model and support vector machine regression model is constructed.(3)After the completion of the model,5 base models,16 ensemble models and one traditional combination model were compared by using DC,USS and MRE as evaluation indexes.The results show that compared with the base model,the ensemble model has a great improvement in three indicators.The DC of the ensemble model are above 0.93;The MRE of DNN-Elman-XGboost model is 19.8%;The USS of DNN-Elman-BP-RF model is 96.9%.(4)At the same time,the four types of ensemble models present different characteristics,such as the XGboost model has a low MRE,while the GBDT model has a high USS.Compared with the traditional combination model,the ensemble model has a more complex construction process,but its accuracy and the fitting of runoff trend are better.