The power spectral density (PSD) function is a frequently used method in the field of stochastic dynamics to determine the governing frequencies of environmental processes, such as earthquakes or wind loads. The PSD function allows buildings and structures to beexamined for stability or is used when planning new buildings. Realistic load models may be generated from real data sets, which can be applied to simulation models. When working with real data records, however, it can be possible that these show a high variance and only few similarities. These differences can often only be detected in the frequency domain, while the data records show high similarities in the time domain. To counteract this problem, this paper proposes a classification approach that determines the spectral similarity of the individual data sets and assigns them to groups using a k-means algorithm. Based on the individual groups, load models can be generated and applied separately to the model to obtain more accurate simulation results. In order to demonstrate the benefits of this classification approach it is applied to numerical examples.