Support Vector Regression for Petroleum Reservoir Production Forecast Considering Geostatistical Realizations

Abstract

Summary Decision analysis related to petroleum field development and management phases with complex models can be time-consuming, especially in highly heterogeneous fields. Probabilistic approaches require a large number of simulation runs to cover all possible solutions, and this can be slow. In this study, we present a methodology to include high-dimensional spatial attributes (geostatistical realizations) in proxy modeling, based on support vector regression (SVR), for building risk curves with decreased run time. The proposed workflow accomplished the following: definition of uncertain inputs, such as porosity and horizontal and vertical permeabilities; selection of outputs from the simulator (cumulative oil, water and gas productions) to be mimicked by the proxies; sample inputs to generate scenarios for training and proxy building; and consistency check to evaluate if the proxy model is reliable to mimic simulator output. We then used proxy models to generate risk curves at the final forecast period (7,305 days) as an application. Using the SVR with high-dimension inputs, we show that the proxy was able to provide reliable results with 300 scenarios, which represent 35% less computational effort compared with using only a reservoir numerical simulator. As a result, we can use this proxy to perform a risk analysis with a high level of accuracy [mean absolute percentage error (MAPE) lower than 0.5%] to predict the production curves. In conclusion, we can use the SVR proxy model technique as an alternative to a reservoir simulator when spatial uncertainty attributes (geostatistical realization) are present.