Prediction of formation pressure in underground gas storage based on data-driven method

Abstract

SUI Gulei 1, , FU Yujiang 1, , ZHU Hongxiang 1, , LI Zunzhao , and WANG Xiaolin 1, 2 1) SINOPEC Dalian Research Institute of Petroleum & Petrochemicals Co. Ltd. , Dalian 116045, Liaoning Province, P. R. China 2) Artificial Intelligence Innovation Center, SINOPEC, Dalian 116045, Liaoning Province, P. R. China Abstract: Formation pressure is a significant parameter for establishing the working system of injector-producer well and monitoring the operation of underground gas storage (UGS). In view of the complex geological modeling and highquality historical fitting involved in numerical simulation to understand the change of formation pressure, a datadriven method for forecasting formation pressure of UGS is proposed. The optimal warping path is weighted by the proportion of gas injection-production to screen pressure monitoring wells. The supervised learning model of forma⁃ tion pressure forecasting is established by three kinds of machine learning algorithms including extreme gradient boosting (XGBoost), support vector regression (SVR), and long short-term memory network (LSTM). The experimen⁃ tal results show that predictive performances of three predictive models are ranked from high to low: SVR, XGBoost, LSTM, among which the predictive performance of SVR is the most stable. Introducing the proportion of gas injection-production to screen pressure monitoring wells can improve the predictive performance of the data-driven model. Research shows that the purely data-driven method can directly interpret routine surface measurements to intuitive subsurface pressure parameters of UGS, which is very suitable for the field application of UGS.