Unlike the extensively researched issue of detecting violent acts in a full video sequence, early violence prediction focuses on recognizing such actions in videos that are only partially available. This is essential for facilitating early interventions in surveillance systems. In this study, we introduce an early violence prediction system that provides high prediction accuracy, even with a limited number of video frames. We introduce an efficient two-stream network which uses 3D ShuffleNet V2 architecture and temporal convolution networks for effectively extracting spatial and temporal features. We integrate early violence prediction by employing a teacher-student framework for training our model. The teacher model, which is trained using full-length videos, distills valuable information to the student model that is limited to using only partial videos. We evaluated our method using the RWF-2000 violence recognition dataset. Our experimental findings showed that the proposed model not only surpasses several advanced violence recognition techniques on full-length videos but also performs exceptionally well in early violence prediction, offering a significant benefit for preventive security measures.