基于聚类算法的支持向量回归建模的新策略

A New Strategy of SVR Modeling Based on Clustering Algorithm

针对支持向量机对时变的样本集采用单一模型建模困难的问题,提出了一种新的学习策略.首先,使用自组织映射(SOM)神经网络和k-m eans聚类算法对初始样本集合进行聚类.然后,针对每个聚类数据集合,通过最优加权组合不同核函数的支持向量回归模型建立最终的模型.实验表明,采用这种学习策略的建模精度要优于单一支持向量回归建模方法.

Aiming at sovling the difficulty of modeling dynamic system with a single model by SVR （ support vector regression）, a new learning strategy is proposed. Firstly, the clustering algorithm combining SOM （ self-organizing map） neural network with k-means algorithm is applied to cluster the original sample set dynamically. Then, the final model of each clustering sample set is established by the optimal weighted combination of different kernel functions of SVR models. The experimental result shows that the proposed learning strategy has much better generalization ability and prediction precision than the single SVR model.