In order to avoid wellbore failure caused by abnormal annulus band pressure and resulting safety accidents, the annulus band pressure value is accurately predicted, and preventive measures are taken in advance when it exceeds the control value. An autoregressive integrated moving average-long short term memory (ARMI-LSTM) model was proposed. The model was trained to predict the annular band pressure of example wells based on actual annular band pressure time series data and feature capture data sets, and compared with a single model and recurrent neural network (RNN) model. The results show that the model has a good performance in error, fitting accuracy and overall performance after training with actual data, which can provide a reference for improving the prediction accuracy and efficiency of annular band pressure value, and is helpful to well integrity design.