Predicting urban cellular network traffic is crucial for urban construction planning and base station construction distribution, but due to a number of complex factors (e.g., dynamic spatial and temporal dependencies, complex spatial dependencies, external environments, etc.), counting the details of these efforts is not easy for current operators. Forecasting for regions is based on historical data to predict future conditions in urban areas. In the previous literature, the more commonly used methods are based on the extraction of temporal features by LSTM and the extraction of local spatial correlation features by local Convolutional Neural Networks (CNN) algorithms for solving region-based problems. On this basis, we take the use of spatio-temporal 3D grouped multiscale ResNet combined with LSTM temporal blocks to achieve the corresponding predictions. ResNet also allows the combination of spatio-temporal features with external factors. Residual Network (ResNet) combines tight and periodic 3DCNN branching with traffic movement trends as well as other external factors to forecast future call-in and call-out traffic for regional cellular network traffic. To assess the effective results of our work, we base it on a real-world cellular network traffic dataset. In contrast to other related current work used for prediction, the experiments demonstrate a reduction in RMSE of 4.7% and 3.3%, respectively, when prediction is performed over a dataset from the Italian city of Milan.