基于密度参数K-均值算法的RBF网络及其在降水量预测中的应用

RBF Neural Network of K-Means Algorithm Based on Density Parameter and the Application to the Rainfall Forecasting

径向基函数(Radial Basis Funtion,简称RBF)神经网络是一种收敛速度快、逼近能力强的前馈型神经网络。为提高网络的训练速度,采用基于密度参数的K-均值算法,消除传统K-均值算法对初始聚类中心的敏感性,构建了基于K-均值算法的RBF降水预报模型,并应用于挠力河流域的友谊农场汛期月降水量预报中,以检验所建模型的有效性。结果表明,与标准的K-均值算法RBF网络模型和BP(Back Propagation)网络模型相比,所构建的RBF降水预报模型对2008年,2009年,2010年各年间汛期(6—9月)降水量的预测平均相对误差为9.270 7%;确定性系数为0.96。预报精度均有所提高,且满足水文预报要求。

The radial basis function （RBF） neural network is a feed-forward artificial neural network with high convergence speed and strong approximation capability. In order to improve the training rate of the RBF, a K-means algorithm based on density parameter was introduced to determine clustering center, which could reduce the sensitivity of traditional K-means algorithm for initial clustering centers. A rainfall forecasting model of RBF based on K-means algorithm was built, which was applied to forecasting monthly rainfall over the Youyi Farm in Naolihe catchment during the flood season, aiming to test the effectiveness of this model. The case study showed that the mean relative error of rainfall forecasting in flood season （from June to September） of the year 2008, 2009 and 2010 was 9. 270 7%, and the deterministic coefficient was 0.96. It demonstrated a higher forecasting accuracy compared to a RBF model based on a standard K-means algorithm and BP （Back Propagation） model, and the rainfall forechsting results met the requirements of hydrologic prediction.