Hybrid Convolutional and Gated Recurrent Unit Network with Attention for Drilling Kick Prediction

Abstract

Summary Drilling safety is a primary issue in the oil drilling process. Kick is one of the most serious accidents in abnormal drilling accidents. If it is not discovered and addressed in time, it may cause a blowout or even a bigger safety accident. Therefore, predicting the occurrence of kicks in advance is very important to avoid more serious accidents. This research introduces a prediction method for kicks using a combination of convolutional neural networks (CNNs) and gated recurrent units (GRUs), along with an attention mechanism, to assess the likelihood of a kick happening downhole in advance. The method uses CNN layers to extract features from drilling data and reduce the dimensionality of these features. It models drilling time series data using GRUs. The output vector from the GRU is weighted by an attention mechanism to focus on more significant features. Finally, the predictions of kicks are derived through data analysis. The results demonstrate that the method can predict the kick 20 minutes in advance with an accuracy of 98.64%. These results will prove to be significant for improving the prediction level of drilling kicks.