Application of Neural Network Parameter Prediction in Reservoir Characterization and Simulation - A Case History: The Rabbit Hills Field

Abstract

Abstract Recent developments in artificial neural networks now make it possible to predict important reservoir parameters from combinations of well log, geologic, and surface seismic data. Our neural network predictions of porosity and permeability make use of high resolution 3-D seismic data and well logs to extrapolate these reservoir properties away from the borehole. These predictions can then be used for further geologic analysis, seismic interpretation, 3-D computer visualization, or input to reservoir simulators. This methodology of integration of data types of different scales combined with prediction of reservoir parameters provides valuable feedback into the reservoir modeling process. This in turn enhances the capability of the reservoir team to model accurately subsurface detail and those features that are likely to control and influence the displacement processes associated with hydrocarbon production. The NE Rabbit Hills Field, north-central Montana, USA, is used to demonstrate the use of this integrated reservoir modeling approach. Results of history matching using the neural network derived rock parameters are also presented. Introduction The petroleum industry has increasingly recognized the need for a cross-disciplinary team approach to reservoir characterization and simulation studies. A reservoir simulation study combines data from many different sources and processes the data through complex nonlinear systems of equations to generate a reservoir production forecast for subsequent economic analysis. The data types typically include geologic, seismic, petrophysical well log, and production data. A significant problem with these simulations is that data generated from different sources are measured at different scales and on different rock volumes. The recent intense focus on reservoir characterization as a production tool has occurred largely because of the enhanced computer capability now available to integrate and visualize large volumes of data in a three-dimensional format. Along with the advances in integration and visualization capabilities, a new modeling method has been introduced that is capable of making parameter predictions in the absence of well defined analytic relationships. This technique, artificial neural networks (ANN), is becoming increasingly popular in the areas of pattern recognition and nonlinear problem solving. Petroleum reservoirs are, in large part, characterized by the porosity and permeability of the hydrocarbon-bearing formation. Accurate estimates of these parameters are important for predicting reserves and developing oil extraction techniques. At present, to obtain porosity or permeability distributions, values are interpolated between existing wells using conventional gridding techniques. Gridding techniques can account for large-scale variations in reservoir parameters but are unable to provide the level of resolution required to produce accurate production estimates or to determine injection locations if the reservoir is located in a zone of complex geology. A neural network approach provides a method for extrapolating well log data guided by another type of information, such as seismic data. Neural networks are thus being used to address the challenge posed by the considerable difference in physical scale of the different data sets that are to be integrated, for example, outcrop and core data with surface and crosswell seismic data. In reservoir characterization, many different types of data must be integrated. Each characterization is limited by the specific data types available and the quality of those data. The objective in a characterization is to provide a three-dimensional model of the reservoir that effectively integrates the available data of different types and scales.