摘要
Gamma Test是一个与模型无关的数据分析方法,可以解决建立回归模型时面对的模型精度评价、划分率定和验证的数据及模型输入因子选择这三个问题。本文以英国的River Tone流域为例,应用Gamma Test方法分析数据,指导建立双层BP神经网络降雨-径流模型。结果表明,Gamma Test可以指导优化输入因子,精简模型结构,防止过拟合。
Nonlinear regression model is widely used in rainfall-runoff modeling. However, estimating the model accuracy, partition data for calibration and validation, and feature selection are three main handicaps when constructing the model. Gamma test is a model-independent data analysis method. It can be used to help constructing model and solve over-fitting problem during model training period. This paper firstly introduced the Gamma Test, and gave the general formula of rain-runoff regression model. Then a case study was taken in the River Tone catchment in the Southeast UK. The error variance estimation by Gamma test provided a target mean squared error for training the two layer back propagation artificial neural network, which was used as the nonlinear rainfall-runoff regression model in this study. Optima input feature combination was selected by the minimum error variance indicated by the Gamma test. The results show that the model with optima feature combination and target mean squared error with error variance perform best during validation period, while model with greater accuracy than the error variance will result in a model overtraining problem.
出处
《水文》
CSCD
北大核心
2010年第1期39-43,共5页
Journal of China Hydrology
基金
国家自然科学基金(50479033)
广东省科技计划基金(2006B37202001)
