Thermal power plant operation depends on the knowledge of a wide range of complex and cross dependentparameters. Information is usually captured through Distributed Control Systems (DCS) which allows to access up todate data but also long periods of recorded operation. Large and available data sets are decisive for plant operation,but they must be properly used and interpreted to achieve effectiveness. The purpose of the present paper is to presentan identification of operational patterns from historical data from an actual thermal power plant based on unsupervisedmachine learning methods. The proposed methodology is applied to a long term data series from the 360 MW Braziliancoal-fired Pecem power plant, for 40 selected parameters, concerning its steam generator and associated mills. Datasetsize and redundancy is treated by the Principal Component Analysis (PCA) approach, which defines a lower dimensionalspace, proper for clustering while preserving most of its variance. The K-means clustering method identifies operatingpoint groups according to their degree of similarity. The appropriate cluster number is defined by means of the averagesilhouette coefficient, which measures the clusters consistency. Cluster parameter values and distribution are evaluatedto verify result consistency. Assessments with the 40 initial parameters and with a subset of 29 parameters from thesteam generator and mills system are presented. The latter’s generate more useful and physically relevant results, beingdescribed globally by a 2 clusters analysis or, for refined observations, by a 10 clusters analysis. The different pat-terns encountered facilitate an understanding of the parameters arrangement and resulting performance, enabling theidentification of higher efficiency operation conditions and supporting practices to improve the plants operation.