随着热工建模过程中参数的增多，根据参数之间的相关性进行分块建模成为降低模型复杂度、提高模型监测效果的有效手段之一。因此提出了一种基于互信息的自动聚类、分块建模方法。首先，获取参数之间的互信息矩阵，在此基础之上以训练数据的平均平方预测误差最小为标准，使用谱聚类算法对参数进行自动聚类。然后，分别建立每个子块对应的主成分分析(Principle component analysis，PCA)模型，并将所有子块的建模结果通过贝叶斯理论进行融合来对多个子块模型进行统一监测。最后，采用基于最小角度回归(Least angle regressions，LARS)的故障诊断方法定位故障发生的方向和幅值。通过数学案例的验证和电厂高温再热器的实际应用，表明了所提方法在故障监测和诊断方面的有效性。

With the increase of parameters in thermal modeling, block modeling based on the correlation between parameters is one of the effective methods to reduce the model complexity and improve the monitoring effect of the model. Therefore, an automatic clustering and block modeling method based on mutual information is proposed. Firstly, the mutual information matrix of parameters is obtained, on this basis,the parameters are automatically clustered by spectral clustering algorithm with the minimum mean square prediction error of the training data as the criterion. Then, the principle component analysis (PCA) modeling corresponding to each subblock is established, and the modeling results of all subblocks are fused by Bayesian theory to monitor the multisubblock models in a unified manner.Finally,a fault diagnosis method based on least angle regressions (LARS) is used to identity the direction and amplitude of the faults. The effectiveness of the proposed method in fault monitoring and diagnosis is demonstrated by verification of mathematical case and practical application of high temperature reheater in power plant.

TK221