A deep learning approach for abnormal pore pressure prediction based on multivariate time series of kick

After a kick occurs during petroleum drilling, the rapid and accurate prediction of abnormal pore pressure is the basis for taking proper well control measures. In this work, we build an end-to-end intelligent model for rapid determination of abnormal pore pressure, which is composed of temporal convolution, graph adaptive learning, and graph convolution. The field kick data of a shale gas reservoir is collected to train and test the model. In the 10 tests, the present model produces a maximum RE of 9.89%, an average RMSE of 0.09, and an average MAPE of 3.9%. An ablation experiment is conduced to evaluate the individual contributions of graph adaptive learning and graph convolution. Compared to the multi-time-step long short-term memory model, the maximum RE is reduced by 93.7%, while RMSE and MAPE are reduced by 82% for both. It is found that the multi-core and multi-length one-dimensional convolutional neural network outperforms the conventional model in extracting multivariate time series features when predicting abnormal pore pressure. Using the strategies of graph structure adaptive learning and graph convolution, the abundance of information and sample diversity in the dataset are greatly enhanced. In general, the present model has high prediction accuracy (96.1%) and reliable robustness in the prediction of abnormal pore pressure, and demonstrates advantages over traditional methods.