Rockburst is an extremely destructive geological disaster in deep underground engineering. In order to accurately predict the intensity level of rockburst, a method for rockburst intensity level prediction based on parallel fusion graph Transformer (PFGT) was proposed. Firstly, the similarity structure relationship of rockburst data in Euclidean space was utilized to construct graph-structured data. Besides, another kind of graph-structured data was constructed by utilizing multiple rockburst criteria to constrain the structural distortion of rockburst data in European space. Single-scale features of rockburst data was obtained through parallel training. Secondly, a feature fusion graph Transformer strategy was designed, which obtains multi-scale features of rockburst data by fusing two types of graph-structured data features based on Euclidean space and based on rockburst criteria. The method improves the data representation capability by simultaneously utilizing single-scale features and multi-scale features. During the training process, using Transformer for feature fusion enables the model to more comprehensively capture the optimized features of rockburst data, thus improving model performance. Compared with traditional neural networks and other machine learning algorithms, the prediction accuracy of the PFGT model is 94.87%, which is superior to other algorithms, proving the effectiveness of this algorithm and providing a new method for rockburst level prediction.