Power Spectral Density(PSD) prediction is an important part of spectrum management. Due to the high complexity, nonlinearity and uncertainty of the PSD, it is difficult for a single prediction model to ensure the accuracy and efficiency of the prediction. In order to overcome the disadvantages of a single prediction method, a hybrid machine learning model is proposed to combine a Self-Organizing Map(SOM) network with a Regression Tree(RT) to predict the PSD of the signal. First, the method uses a self-organizing map network to cluster the original sample sets with similar manual features. Then, a RT is constructed for each cluster to predict the PSD. Finally, the data of RWTH from Aachen University are adopted for experiments. The root mean square error of the prediction result is 0.824 higher than that of the existing method, which proves that the hybrid model has higher prediction accuracy and better generalization ability.