With the frequent occurrence of social violence, predicting violent behavior becomes the key to intervention. In this study, a categorical prediction method of violent tendency based on scalp EEG data was proposed. First, the EEG data of the subjects were collected, and then filtered and independent component analysis (ICA) was carried out to extract relatively clean EEG signals. Then, the power spectral density features of the processed EEG signals are extracted. A hybrid neural network is then designed, combining convolutional neural network (CNN), multi-head attention mechanism (MHA) and long short-term memory network (LSTM), as well as activation layer, exit layer and fully connected layer. The PSD features of EEG signals are used as inputs for training hybrid neural networks. The experimental results show that the accuracy of the model is 92.36%, the accuracy is 92.43%, the sensitivity is 97.15%, the specificity is 80.82%, the F1 score is 94.73%, and the training time is 18.21 s, which is better than the mainstream neural network. In addition, psychological data was collected, and a feature-level fusion method was used to fuse EEG data with psychological data to form a complete feature set, which was input into the CNN-MHA-LSTM network for training. Accuracy 97.99%, accuracy 97.76%, sensitivity 99.43%, specificity 94.52%. F1 scored 98.59%. Compared with the use of EEG data alone, the performance is significantly improved, which proves the potential and advantage of the integration of EEG data and psychological data to predict violent behavior.