The Investigation On Optimal Estimation For Offshore Gas Production System Based On Inverse Problem Solving

With the development of offshore oil and gas fields,the monitoring of corresponding production systems are faced with severe challenges: complex marine environment,high cost of installation and maintenance for multiphase flowmeters and interference of various errors on the temperature and pressure measurements of lead to difficulty in obtaining reliable operation information,preventing safety production.To overcome such problems,based on the idea of inverse problem solving,this paper carries out the studies of process identification and soft sensing for offshore natural gas production system with data-driven and model-driven optimal estimation algorithms established.These established algorithms are computer-based soft sensing solution,which are intended to promote the soft sensing module of flow management system,improve its data acquisition ability and provide online monitoring tool for central control technicians as well.For the simultaneous estimation of multi-well flowrates in gas production system,a model-driven static estimation algorithm is built based on data reconciliation technique.The proposed algorithm is intended to search for optimal flowrates estimation based on weighted least square criterion,comprehensively considering measurement and model errors,integrating two-phase physical flow models of multiple flow units as constraints.Hybrid parallel genetic algorithm is introduced as the optimization solution,effectively covering the defects of time-consuming evolution of simple genetic algorithm and improving the calculation efficiency.The static estimation algorithm is validated by the field data of a realistic gas production system.The results indicate that the proposed algorithm maintains high accuracy and robustness with absence of individual pressure sensor or total flowmeter.In addition,the adopted optimization method shows good parallel performance,where the computing time can meet the needs of engineering applications.Based on dynamic training data,a set of model bank made up of polynomial NARX and DNN-NARX black-box models is built for the large time-delay gas production system,using orthogonal least square regression and deep learning technique,approximately describing the dynamic performance of each well,in order to calculate single-well flowrate and pressure.Subsequently,with static training data,the parameters of above black-box models are respectively corrected by bi-objective least square algorithm and transfer learning technique,improving global applicability of model bank.Through realistic operation data,the simulation results of polynomial NARX model,DNN-NARX model and multiple-layer-perception(MLP)-NARX model are compared.It is observed that DNN-NARX model shows the best performance with advantages of higher accuracy,better approximation ability and stronger generalization ability.In addition,superior identification method and model settings with strong expansibility and engineering practicability are recommended for soft sensing problems in the petroleum industry.Combined with proposed NARX model bank,a data-driven estimation algorithm is established by unscented Kalman filter(UKF),in order to realize closed-loop online estimation of single-well flowrate and pressure,making up for the accuracy limitation of NARX model.Firstly,the state space equation of each well is built based on a NARX model,with which the estimation unit,UKF,is constructed;due to the redundancy of NARX models,two filter fusion methods(FF1 and FF2)are established to improve the performance of soft sensing algorithm.Through realistic operation data,the estimation accuracy as well as calculation cost of single UKF,FF1 and FF2 are compared.Results indicate that,FF1 shows highest global accuracy and relative high calculation speed and is consequently recommended as the preferred data-driven soft sensing scheme.The single UKF constructed by DNN-NARX grey-box model is recommended as an alternative with least time consuming and second highest accuracy.On the basis of steady state physical model(WST)of gas-liquid two-phase flow corrected by time-varying coefficients,a set of dynamic model-driven estimation algorithm(WST-AUKF)is built introducing augmented unscented Kalman filter(AUKF),covering the shortage of limited applicability in data-driven algorithm.The time-varying coefficients,as augmented state variables,are estimated simultaneously with flow rate and pressures,used for real-time correction of physical model,further improving its adaptation to actual working conditions.The feasibility,accuracy and online applicability of WST-AUKF are verified by semi-realistic operation data.Results indicate that,the WST-AUKF can accurately estimate state variables and effectively correct model parameters in the meantime.Its calculation time can meet the demand of online estimation in engineering application.