Research On Data-driven Approaches For Production Prediction Of SAGD Process

China has rich heavy oil reserves which are,however,difficult to be exploited due to high viscosity and poor fluidity.Steam Assist Gravity Drainage(SAGD),a kind of thermal technology,enhances oil recovery rate prominently by steam viscosity reduction,which has good production effects on oil fields.It has significant practical implications to establish a SAGD production prediction model for purpose of development scheme formulation and timely production scheduling execution of oil field,and relevant researches have been carried out abroad.Nevertheless,abroad SAGD modeling approaches are not appropriate for domestic production process on account of diverse well pattern deployment.In this thesis,research on domestic SAGD production prediction,simulations and analyses are undertaken based on above issues.Main contributions of this paper are summarized as follows.(1)Typical characteristics of domestic SAGD,including several injection wells and production wells as well as intercoupling among production wells,injection wells and the above two types of wells,are drawn from the well pattern deployment at home.A research is carried out about three modeling approaches designed for abroad SAGD production features.Model performances are verified based on the real-life industry production data from certain domestic oil field and demonstrate the lack of fitness of three approaches.A SAGD production prediction model based on time-delay dynamic partial least squares is constructed on account of the features of multivariable coupling,time-delay and dynamic of domestic SAGD.Comparing the simulation results of the real-life production data with those of the existed approaches,it interprets the effectiveness of model performance improvement when taking time-delay and dynamic features into consideration,and the superiority of the purposed approach.(2)To tackle the poor model generalization performance reliability of the data-driven model,an online model switch strategy is proposed based on time and space similarity principle in view of approach(1)above.The proposed approach constructs an offline model set,tests the comprehensive generalization ability for each candidate model,and selects the optimal model to estimate production.Comparing the proposed strategy with approach(1)above,it specifies the outstanding availability of the compound principle.The simulation results of the real-life production data illustrates that the proposed approach is able to switch model online precisely,improves model prediction accuracy and addresses model generalization performance reliability successfully.(3)An adaptive model switch strategy is proposed on account of the slow time-varying feature of SAGD,which implements online model updating,selection and switch and predicts production combined with moving window technique.The simulation results illustrate the proposed approach adapts the slow time-varying feature of SAGD more precisely and improve further model prediction accuracy.