关键词: 径流预报;时变权重;随机森林;BP神经网络;多元线性回归;北碚水文站;嘉陵江流域;

Abstract: In order to improve the accuracy of runoff forecasting, taking Beibei Station in Jialing River as an example, the runoff data from 1979 to 2019 were selected to compare and analyze the forecast performances of various modeling methods for monthly runoff forecasting. Time series method was selected as a typical method for univariate forecasting model, random forest, BP neural network and multiple linear regression were selected as the typical methods for multivariate forecasting model, and time-varying weight was applied for combined forecasting. By comparing the forecast results of univariate and multivariate, single and combined models, the forecast performances of the mentioned modeling methods was analyzed. The results showed that: (1) The multivariate forecast models showed better forecast performance and the average absolute percentage error was reduced by about 20% compared with the time series model. (2) Two machine learning methods, random forest and BP neural network, had better predictive ability in single models. (3) The time-varying weight combination forecast method could effectively combine the advantages of single models to improve the forecast accuracy. The research results can provide a reference for the formulation of modeling strategies for medium and long-term runoff forecasting.

Key words: runoff forecasting; time-varying weight; hydrologic series; random forest; BP neural network; multiple linear regression; Beibei Station; Jialing River