Thermal power units have been widely put into operation for the electrical peakshaving, which results in the increase of unsteady state operating conditions and the deviation of common operating conditions from design conditions. Thus, the operating condition classification model based on the historical data clustering is proposed in this work. Firstly, considering the coexistence of unsteady and steady state operating conditions, the output power is applied as the key indicator between the steady state and unsteady state. The interval estimation of expectation of the output power difference value is used to classify the historical data into the steady and unsteady samples. Then, due to the distribution difference among external boundary variables under the steadystate operating conditions, the improved multistep Kmeans clustering algorithm is proposed. The optimal clustering number for each step is determined by using the silhouette evaluation criterion. Finally, a real heavy gas turbine is used to validate the established model. Compared with the traditional Kmeans clustering, the results prove that the proposed operating condition classification model can effectively solve the problems of less classifications of operating condition and uneven distribution of samples.