Dynamic Structure Learning of Factor Graphs and Parameter Estimation of a Constrained Nonlinear Predictive Model for Oilfield Optimization.

摘要: Injector-producer relationships (IPRs) are the key knowledge for oilfield optimization, i.e., maximizing oil production at minimum operational cost. The difficulty associated with field optimization is that underlying reservoir structure unknown and changes continuously over time. Inferring IPRs a large-scale constrained nonlinear parameter estimation problem. state-of-the-art hybrid (HCNO) method provides excellent accuracy solving this problem but prohibitive computational costs large oilfields. In paper, we propose dynamic learning approach based on inference in probabilistic graphical model named PETROGRAPH Learning (PGL). initiated by constructing an initial factor graph locality principle guided belief discrepancies, error, residual correlation analysis. At each iteration, sum-product algorithm applied to estimate parameters, refined as input next round. iterative continues until convergence. Experimental results analysis show PGL scalable scale of real oilfields much less running time than HCNO while providing virtually exact solutions.