Taking thermal efficiency and LOI as outputs and 20 operating parameters as inputs of the combustion process of a 600 MWe cyclone boiler with 6 mills, 10 fuzzy rules are set up by K-means clustering firstly according to the data collected by DCS. Secondly, a T-S (Takagi-Sugeno) fuzzy neural network which consists of 11 BP sub-networks is developed to model the combustion process characteristics. As the LOI measuring device does not work well and thermal efficiency is a calculation value which is dependent on other primary parameters especially those of coal quality which are hard to be measured accurately and timely, the data collected are unavoidably spoiled by noise seriously. Normal neural networks such as BP and RBF are not capable of dealing with such a noise case. Validation tests demonstrate that the proposed T-S fuzzy neural network, which synthesizes well the advantages both of neural network and fuzzy reasoning, is much friendly to data noise and can be chose as an applied data mining tool.