Predictive process monitoring has become a major contributor to data-driven insight in process mining. Process prediction, which is used to extract models from historical event logs to predict trace evolution, has become one of the main driving forces in process mining. However, the unequal distribution of data in event logs affects the final prediction accuracy. This paper proposes a process mining method, which employs the structure of Convolutional Neural Network (CNN) combined with Long Short-Term Memory (LSTM) to provide a method to predict the next activity in the running trace for the execution scenarios of business processes. Firstly, event logs containing historical data are converted to image storage. Then, the spatial image data are feature engineered and eventually used to train the designed CNN-LSTM network to learn a deep learning model that can predict the next activity. The proposed framework improves the accuracy and training performance of the original method (reducing the influence of overfitting). Two event logs are used to test the feasibility of the proposed framework, and it is proved that the proposed framework has higher accuracy than the previous network.