The present study proposes a novel single-layer Neural Network proxy to efficiently predict the production performance of oil reservoirs from a limited number of reservoir simulations. The proposed model is shown to provide powerful means for learning reservoir's dynamics from input-output relationships that is defined by multiple combinations of inputs and controls. A SAGD case with 3 well pairs is used to illustrate the approach. The workflow is organized as follows:

  1. Different numbers (from 30 to 200) of direct numerical simulations are conducted for the time horizon of ten years for the given reservoir. These simulations correspond to different combinations of the operational parameters sampled according to the Latin Hypercube Experimental Design (LHD).

  2. The time series of the simulated entire field production performance (particularly: Cumulative Oil Production, cumulative SOR, Oil Recovery Factor, and Oil Production Rate) are used to build the corresponding Radial Basis Function (RBF) Network proxy model which represents production performance as objective functions of the operational parameters and time. The nodal Radial Functions of the Network are defined to exactly match the simulator's training outputs. The big advantage of this model is a combination of low computational complexity with high prediction accuracy.

  3. The proxy model is then used to predict the production data for the given reservoir for any time period (within 10 training years) and any combination of the operational parameters.

  4. The predicted data are compared with the actual simulation results for the same time period and the same combinations of the operational parameters to evaluate the prediction quality. It is shown that the proposed RBF proxy model can be used as a light version of simulator to estimate the production results for any combination of operational parameters and time horizon with much less computational efforts. The proposed approach is shown to provide a very efficient forecasting mechanism for the reservoir considered. The difference between the predicted and actual data could be as low as a few percent for the majority of the operational parameters depending on the number of simulations used to train the RBF model. In general, the accuracy of the proxy model increases with the number of training simulations.

Keywords:
thermal method, SAGD, Artificial Intelligence, machine learning, neural network, reservoir simulation, proxy, production rate, Radial basis Function, simulator
Subjects:
Improved and Enhanced Recovery, Reservoir Simulation, Information Management and Systems, Thermal methods
Copyright 2014, Society of Petroleum Engineers
You can access this article if you purchase or spend a download.