Downhole quantitative evaluation of gas kick during deepwater drilling with deep learning using pilot-scale rig data

Abstract Gas kick occurs frequently during deep-water drilling operations caused by the lack of safe margin between pore pressure and leakage pressure. The existing research is limited to gas kick classification and cannot quantitatively evaluate the gas kick risk in the downhole very well. Thus, the objective of this work is to systematically use Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN) models based on pilot-scale rig data for quantitative evaluation of gas kick risk. Furthermore, the quantitative evaluation is not surface but downhole. First, the gas kick simulation experiment is accomplished in the pilot-scale test well and produces the gas kick dataset, which is based on the multi-source data fusion through the surface monitoring technologies, riser monitoring technologies and downhole monitoring technologies. Second, the training features are selected and grouped as Sets1-5 to study the features' sensitivity. Third, the raw data is processed and prepared for the following machine learning framework. Fourth, there are five (5) LSTM models trained on Sets1-5. The results indicate that the models’ Loss decrease with the increase of feature number, which has fully demonstrated the effectiveness of PWD, EKD, and Doppler parameters. Finally, there are four representative case studies (artificial gas kick) that are used to test the above five models. The compressed air injected rate (AR) prediction error and detection time-delay decrease with the increase of feature number. The LSTM model trained with the combination of surface-riser-downhole comprehensive detection technologies performs the best in reducing both the prediction error and detection time delay, which could be used to quantitatively evaluate the downhole gas kick risk in the more accurate, faster, more stable, more reliable, and cost-effective manner, and it is effective and worthy of promotion.