Air-conditioning system operation pattern recognition plays an important role in the fault diagnosis and energy saving of the building. Most machine learning methods need labeled data to train the model. However, the difficulty of obtaining labeled data is much greater than that of unlabeled data. Therefore, unsupervised clustering models are proposed to study the operation pattern recognition of the refrigeration, heating and hot water combined air-conditioning (RHHAC) system. Clustering methods selected in this study include K-means, Gaussian mixture model clustering (GMMC) and spectral clustering. Further, correlation analysis is used to eliminate the redundant characteristic variables of the clustering model. The operating data of the RHHAC system are used to evaluate the performance of proposed clustering models. The results show that clustering models, after removing redundant variables by correlation analysis, can also identify the defrosting operation mode. Moreover, for the GMMC model, the running time is reduced from 27.80 s to 10.04 s when the clustering number is 5. The clustering performance of the original feature set model is the best when the number of clusters of the spectral clustering model is two and three. The clustering hit rate is 98.99%, the clustering error rate is 0.58% and the accuracy is 99.42%.