(Begin page 1893)
AAPG Bulletin, V.
Detection of hydrocarbon reservoir boundaries using neural network
analysis of surface geochemical data
1Global Customer Solutions Management, i2 Technologies, 600
Lanidex Plaza, Parsippany, New Jersey, 07054; email: [email protected]
2Pennsylvania State University, Hazelton, Highacres, Hazelton, Pennsylvania, 18201; email: [email protected]
3Energy and Geoenvironmental Engineering Department, Pennsylvania State University, 152 Hosler Bldg, University Park, Pennsylvania, 16802; email: [email protected]
Hari Doraisamy holds a B.S. degree in chemical engineering from Biria Institute of Technology and Science, India, and an M.S. in petroleum and natural gas engineering from the Pennsylvania State University. He worked with the Dowell division of Schlumberger for a year and a half and joined i2 Technologies, Inc. as a demand planning consulting. His main area of interest is in artificial intelligence, particularly for forecasting various events. He has presented papers involving the use of neural networks in the area of reservoir simulation at various Society of Petroleum Engineers conferences.
Daniel H. Vice has a B.S. degree in geology from Oregon State and an M.S. degree in geology from Arizona State. He worked for the U.S. Bureau of Mines in 1971, locating, mapping, and sampling old mines and claims in the Idaho Primitive Area (now the Frank Church Wilderness Area). He also worked for Burlington Northern from 1971 to 1982, primarily doing geothermal exploration in the Washington Cascades. Vice completed an D.Ed. in earth sciences from Penn State in 1996. Since 1997 he has been teaching at Penn State Hazleton and Penn State Schuylkill. His current research interest is surface geochemical sampling methods for oil and natural gas exploration.
Phillip M. Halleck is an associate professor of petroleum and natural gas engineering. He holds a B.S. degree in chemistry from the University of Rochester and a Ph.D. in geophysics from the University of Chicago. He has a wide range of experience in explosives and geotechnical applications, including rock physics, coal mine subsidence, methane hydrates, well perforating and completions, and sand production. He teaches courses in energy and the environment, well logging, and well completions.
We gratefully acknowledge the cooperation of Consolidated Natural Gas Transmission Corporation and the landowners along the survey line for permission to collect samples. We appreciate the help of Scott Heckman and Dan Ottenstein in performing the gas chromatographic analyses. R. Watson, A. Grader, and T. Ertekin provided critical review and discussion of the manuscript.
Surface geochemical surveys could become important tools for defining the boundaries of a hydrocarbon reservoir. Conventional statistical analysis has shown that a correlation can indeed be found between surface geochemical data and the location of a sample site with respect to the boundaries of a known reservoir. However, such analysis methods cannot be used directly as predictive tools. This article describes the successful application of artificial intelligence in the form of neural network analysis to determine whether a specific sample site, given the ethane concentration in the soil and certain environmental data, is within the surface trace of the reservoir boundaries.
Data from a previous study over a known gas storage reservoir were used to train a back-propagation neural network. No attempt was made to optimize the structure of the network. We used 85% of the data to train the network and withheld 15% to act as unknowns. The input variables consisted of adsorbed ethane concentration and a series of soil description and environmental parameters. The output variable was a simple binary reflecting whether the sample site was directly over the reservoir. The final network was able to predict 95% of unknown sample sites. We found it necessary to include in the input data the ethane concentrations for sites on either side of each site studied. This is consistent with previous observations that a series of adjacent sites having anomalous concentrations hold more significance than do isolated sites. We also found that the use of the land (probably reflecting the degree of disturbance) and soil moisture are the most important environmental variables. This is consistent with previous conventional statistical studies of the same data. We conclude that application of neural networks to properly designed surface geochemical studies holds promise for use in defining the boundaries of known reservoirs.
(Begin page 1894)