基于MI-PCA与BP神经网络的石羊河流域中长期径流预报

Long-term Runoff Forecasting Model Based on MI-PCA and BP Neural Network in Shiyang River Basin

使用神经网络进行水文预报的关键之一是预报因子(输入变量)的筛选。鉴于现有方法对预报中因子与径流间复杂的非线性相关关系考虑不充分以及因子间信息重叠导致的算法“过拟合”等问题,提出了一种信息熵理论和主成分分析方法相结合的预报因子筛选方法,并应用至石羊河流域的中长期径流预报中。实例研究表明:运用基于MI-PCA的预报因子筛选方法构建的石羊河流域中长期径流预报BP神经网络模型检验期预报合格率为91.67%,优于单独基于互信息法(83.33%)和主成分分析法(75.00%)的合格率,预报精度满足相关标准规范的要求,可为石羊河流域中长期径流预报提供实际支撑。

Selecting predictors (input variables) is the key for hydrological forecasting based on neural networks. Aiming at the status quo that some existing methods cannot fully reflect the complex nonlinear relationship between predictors and predicted runoff and the information overlapping of the predictors can easily lead to overfitting of the forecasting model, this paper proposes a predictor screening method combining information entropy theory and principal component analysis and applies it to the long-term runoff forecast of the Shiyang River Basin. The case study shows that the qualified rate of the long-term runoff forecast model based on the artificial neural networks in the Shiyang River Basin constructed by MI-PCA method is 91.67%, which is better than that based on the mutual information method (83.33%) or the principal component analysis (75.00%) alone. Therefore, the forecast accuracy meets the requirements of relevant standard specifications and can be used as a supporting model for long-term runoff forecast in the Shiyang River Basin.