Study on Different Integrated Learning Models for Predicting Short-Term Inflow Runoff in Small Watershed Based on Various Meteorological Elements
Zhu Zhanyun
Chen Guangyu
Si Wei
Zhang Wenwen
Zhang Weiwei
Abstract：Improving the accuracy of short-term runoff forecast model is important for the decision-making of reservoir operation.Based on the meteorological and hydrological observation data of Zeya Res-ervoir watershed,four integrated learning models of random forest,GBDT,AdaBoost,and Stacking are used to simulate daily runoff forecasting in small watersheds,and the influence of meteorological factors such as temperature,pressure,and humidity on runoff prediction and simulation is comparatively ana-lyzed.The results show that:(1)The random forest,GBDT,and Stacking models based on multiple meteorological elements have good simulation and prediction capabilities for short-term inflow runoff in small watersheds.The compliance rates of these models during the verification period exceed 70％,meet-ing the accuracy requirements for issuing formal forecasts.Among them,GBDT is the best in terms of prediction and measurement and random forest is the best in terms of accuracy.Therefore,they can be comprehensively applied for decision-making in actual operations.(2)The introduction of meteorological elements(temperature,pressure,and humidity)can significantly improve the accuracy of short-term runoff prediction by each integrated learning model.In particular,the compliance rate of base and medi-um flow prediction by random forest,GBDT,and Stacking models can increase from about 40％to more than 80％,the efficiency coefficient increase to about 0.85,while the average absolute error can decrease by about 50％and the average relative error decrease by about 60％.These findings have high reference value for runoff prediction and simulation during non-flood periods.
Keywords：meteorological elementintegrated learninggradient boostingrandom forestrunoff forecast
Publication Date：2025-07-30
Online Publishing Date：2026-09-12(First online date of this platform, not the publication date of the document)
Pages：9( 62-70 )