Efficient Dimensionality Reduction Methods in Reservoir History Matching

Production forecasting is the basis for decision making in oil and gas industry, can be quite challenging, especially terms of complex geological modeling subsurface. To help solve this problem, assisted history matching built on ensemble-based analysis such as ensemble smoother Kalman filter useful estimating models that preserve realism have predictive capabilities. These methods tend, however, to computationally demanding, they require a large size stable convergence. In paper, we propose novel method uncertainty quantification reservoir model calibration with much-reduced computation time. This approach based sequential combination nonlinear dimensionality reduction techniques: t-distributed stochastic neighbor embedding or Gaussian process latent variable clustering K-means, along data assimilation multiple assimilation. The cluster used reduce number initial geostatistical realizations select set optimal similar production performance reference model. We then apply providing reliable results. Experimental results Brugge field case verify efficiency proposed approach.