基于改进BP网络的日径流预报模型研究

Study on Daily Runoff Forecasting Based on Improved BP Network

针对常规BP算法收敛速度慢和难以获得全局最优的不足,将网络误差函数的改变量引入权值和偏移值的调整,采用自适应学习速率和自适应动量因子调整策略,建立了基于多层感知器神经网络(MLP—ANN)的水文预报模型。采用自相关函数(ACF)和交叉相关函数(CCF)确定网络输入因子并使用试错法优化网络结构。以湖南省双牌水库日入库流量预测为应用实例,并将模拟结果与常规BP网络模型和新安江模型进行对比分析。结果表明,改进模型收敛速度快、预报精度高。

According to the slow learning convergence and local minimum of conventional BP （Back Propagation） algorithm, an improved learning strategy combining the error function variation with the adjustment of weights matrix and biases matrix was proposed and a multilayer perceptron artificial neural networks （MLP-ANN） model based on this self-adaptive algorithm was developed for hydrological forecasting. The auto-correlation function （ACF） and cross-correlation function （CCF） analysis was used to determine the predictors and the trial-and-error measure was taken to optimize the network structure. The Shuangpai reservoir in Hunan province was selected as an example to demonstrate the modeling and the forecasting result was compared with that of conventional BP model and Xin＇ anjiang model. The experimental results show that the improved BP model is much more efficient in time saving and global optimization, and the forecasting accuracy is much better than that of the other two models.