Abstract: Fluid injection-induced seismicity threatens the safe operation of oil and gas, geothermal, and carbon-storage projects. Accurate prediction of its key characteristics can reduce seismic risk. Previous studies have made progress in forecasting magnitude, timing, and seismicity rate, but research on hazardous-zone prediction remains limited. We propose a seismic catalog processing method based on the spatiotemporal distribution feature engineering of seismic activity intensity. A hybrid CNN-LSTM model is then established, combining the spatial-feature extraction capability of convolutional neural networks with the temporal sequence modeling of long short-term memory networks. Model performance is assessed using four metrics: hazardous-zone accuracy, seismic activity intensity error, and accuracy and error of peak-intensity zone prediction. The results are compared with those from LSTM, ConvLSTM, and linear regression model. Results show that the proposed spatiotemporal distribution feature engineering of seismic activity intensity simplifies the modeling process while preserving essential physical information such as event time, location, and intensity. The CNN-LSTM model achieves an accuracies of 72.74% and a mean squared error of 0.22 for seismic activity intensity prediction on the test set. Compared with the other three models, the CNN-LSTM model can more accurately extract the spatial distribution characteristics of seismic activity intensity and performs best in all four performance metrics. The spatiotemporal evolution patterns of hazardous zones learned by the CNN-LSTM model are statistically consistent with the underlying physical mechanisms governing induced seismicity. These findings demonstrate that integrating spatial and temporal information provides a more reliable approach for forecasting seismic hazardous zones during fluid injection.