梯级-关联算法在多泥沙河流含沙量预报中的应用

Application of cascade-correlation algorithm to forecast sediment concentration in heavily silt-carrying rivers

本文借助历史加成法处理样本数据,并分别利用梯级-关联算法(CC)和误差反馈传播算法(BP)建立模型对黄河下游夹河滩水文站汛期含沙量进行预报。传统BP网络需要预先设定网络结构,预报过程虽利用了神经网络的内插特性,但其样本的处理方式和网络构建方式使得运算效率较低;CC算法仅要求初始网络含有输入层和输出层,通过运算不断向网络增加隐含节点,从而最大限度的减少了在网络构建过程中的主观因素。本文比较了当预报的峰值超出训练样本取值范围时两种算法的表现,结果显示:当预报的峰值为训练样本峰值的2.45倍时,二者均能实现较为准确的预报,BP网络在预报精度上要略高于CC网络,但CC网络在运算速度上要明显快于BP网络。

The models based on network algorithm for forecasting cascade-correlation （CC） algorithm and back-propagate （BP） neural the sediment concentration at Jiahetan Hydrological Station in the Lower Yellow River during flood season are developed based on observation data. The samples of data for training are standardized in a pattern of adding an increment to the observed historical value. It is found that the standard BP algorithm has the advantage of good in interpolation, but the structure of the network must be predefined which resulted in low calculation efficiency. Whereas, the initial network structure of CC algorithm is merely composed of input layer and output layer and the hidden neural cells are inserted into the network one by one in the training process. The comparison of the forecasting values beyond the range of training data shows that both methods can realize high accuracy forecasting even the value is higher than 2.45 times of the training peak value. It is concluded that the performance of BP model is a little bit better than the CC model, while the calculation speed of CC model is much higher than that of BP algorithm with better forecasting accuracy.