Abstract: Objective: Crime forecasting can predict the probability and trend of criminal activities in a region, optimize the allocation of law enforcement resources, reduce crime rates and improve social security. Because of the changing laws of spatial similarity, periodicity and spatial heterogeneity of crime events, existing crime prediction methods give less consideration to the characteristics of spatial similarity in crime regions and have strong limitations in facing the problem of long-term spatial and temporal dependence on spatial and temporal data, resulting in limited prediction effects. Methods: The model is divided into two parts: spatial similarity feature capture and prediction, which are composed of a regional spatial similarity feature capture network and a feature fusion coding-based crime prediction network, respectively. The model utilizes the spatial channel attention mechanism and designs a spatial channel attention multi-graph convolutional network based on a multi-graph convolutional network to achieve spatial similarity feature capture. On this basis, the crime event information is represented by embeddable linear sequences, and then the captured spatial similarity features of the crime region are fused to establish a spatio temporal crime feature representation coding; in order to enhance the long spatio temporal dependence of the prediction model, a Transformer prediction network based on the multi-head spatio temporal attention mechanism is designed. To validate the effectiveness of the proposed model, the paper conducts experiments with crime event data from different periods in Chicago and Los Angeles areas and compares it with STGCN, ST-ResNet and ConvLSTM methods. Results: The experimental results show that the proposed model not only outperforms other models on large scales but also exhibits stronger accuracy and stability on small scales in different regions. Conclutions: The methodology improves the accuracy of crime prediction.