Using Artificial Intelligence to Predict Drilling Success Using Regional Data, Brushy Canyon Formation, Delaware Basin, New Mexico
نویسنده
چکیده
Incomplete and sparse information can introduce a high level of risk for oil exploration and development projects. “Expert” systems developed and used in several disciplines and industries have demonstrated beneficial results in modeling the decision making process of knowledgeable experts. A state-of-the-art exploration “expert” tool using a computerized data base and computer maps generated by neural networks, is being developed using fuzzy logic, a relatively new mathematical treatment of imprecise or non-explicit parameters. Analysis to date includes generation of regional scale maps of aeromagnetic, gravity, structure, thickness, and production data for the target Brushy Canyon Formation in the Delaware Basin, New Mexico. For each regional scale map, data attributes were also computed to look for more subtle trends. These attributes include directional first and second derivatives, dip azimuth and magnitude, and azimuth and magnitude of curvature. These data were mapped and gridded to a 40 acre spacing, the current well spacing for Delaware pools in New Mexico, and compared to average monthly production in the first year for Delaware Brushy Canyon wells. The geophysical and geologic data covers 60478 bins (3780 square miles), of which 2434 of these bins have oil, gas and water production data. Using a new fuzzy ranking tool each data attribute was ranked for its ability to predict production potential at these well locations. The highest ranked attributes were gravity dip-azimuth, second latitude derivative of thickness, longitude derivative of gravity, and longitude derivative of structure. These attributes were used to generate a production potential map for the Delaware basin, using neural networks and expert systems, at the scale of 40 acres. Expert systems operate by developing rule sets that can be used to answer questions related to the problem at hand, in this case prospect evaluation. Since prospect generation data often contains non-crisp data, such as “low porosity” or “high on structure”, the expert system will necessarily allow fuzzy inputs. The approach taken is to accumulate all available public domain data and incorporate them into online databases, which can be accessed by the expert system. The primary goal after development of the production potential map is to teach the expert system to add or detract to each prospect’s estimate of risk, in a fashion similar to that employed by human explorationists. Such a map would be a useful tool for evaluating the potential of infill, step out, and wildcat wells in the Delaware basin, both at reservoir and regional scales. This paper discusses the development of this production potential map and initial rules development for the expert system. 2 R. S. BALCH, D. M. HART, W. W. WEISS, AND Ft. F. BROADHEAD 245 Introduction Expert systems are computer programs that are designed to make decisions similar to the manner in which a human expert would. In the past expert systems have been primarily restricted to medical and industrial applications, but with DOE support an expert system to prospect for oil is now being developed to automate and accelerate prospect development for the Brushy Canyon formation in the Delaware basin. Expert systems operate by developing rule sets that can be used to answer questions related to the problem at hand, in this case prospect evaluation. Since prospect generation data often contains non-crisp data, such as “low porosity” or “high on structure”, the expert system will necessarily allow fuzzy inputs. The approach taken is to accumulate all available public domain data and incorporate them into online databases, which can be accessed by the expert system. A primary goal was to develop a map of production potential based on available regional data from which the expert system could add or detract to each prospects estimate of risk. Regional Data A key component to the success of this study is the analysis of the regional data to provide baseline data to correlate with production potential; this also provides a source of heuristic rules for the expert system. Four major categories of regional data were selected and compiled. Regional gravity surveys cover the entire area of the Delaware basin and have been compiled with an accuracy of a few milligals. The survey measurements are on the order of a few thousand feet apart, but sample point locations are highly variable as gravity is measured in easily benchmarked locations, such as along roadways. Gravity measures variations in density and tends to highlight large-scale regional structures at basement depths and if structure has an impact on maturation, migration or trapping of hydrocarbons in the basin useful information can be obtained. Regional Aeromagnetic data, primarily collected via over-flights with 1 mile spacing re-gridded to 0.296 miles longitude and 0.346 miles latitude, also exists for the region. Aeromagnetic data highlights contrasts in the magnetic susceptibility between rocks and can help indicate basement blocks, large-scale faults, and possible large-scale alluvial deposits. The structure of the lower Brushy Canyon was picked on 700 wells in the basin covering a geographically large area’. Large-scale maps of these attributes covering the region were constructed with a kriging algorithm. Structure can play more than one role in trapping and migration of hydrocarbons and two potentially helpful attributes for this study are structural highs, and flexures, which while commonly induce fracturing along the flanks of structures may also help locate subsurface faults which can compartmentalize fields, or allow preferential migration paths for hydrocarbons and water. Finally, the wells used to compute structure were used to generate an isopach map for the Brushy Canyon in the region(Broa&ead and Justman, 2000). Thickness may indicate areas of greater potential production and also can indicate pinch-outs and other nonstructural features that may form hydrocarbon migration pathways or traps. A number of attributes were calculated from these 4 principal data types. These attributes are 1”’ and 2nd derivatives along latitude and longitude; dip azimuth and magnitudes; and curvature azimuths and magnitudes. These values were computed to expose finer scale features in the basic data types that might be useful for correlating back to a production indicator. A total of 36 maps were generated using the Zmap tool of Landmark Graphics Release 98 plus interpretation package (Table 1). Each of these maps was gridded at a scale of 1320-t? (quarter section) because that is the regulatory spacing for wells in the Brushy Canyon in New Mexico. The gridded data was 3 R. S. BALCH, D. M. HART, W. W. WEISS, AND R. F. BROADHEAD 246 exported and loaded into the project production database. Our current production database is a subset of the Onguard database (courtesy of the SW PTTC) containing production information on all New Mexico wells. In this database we have also identified Brushy Canyon wells and using grid locations from the Zmap maps we were able to correlate producing wells with grid numbers. This essentially allows regressions to be formed using the production data as control points (training and testing) and the attribute data as variables. Any regression formed in this manner could then be used to predict production in all other 60,478 40-ac bins in the basin. Fuzzy Ranking There are two primary considerations when trying to form regressions: the first involves the quality of the data you are attempting to predict with the generated regression model; the second deals with the choice of attributes or variables that will be used in forming the regression model. An optional consideration is the application of linear models (least squares regression) or more complicated non-linear solutions such as polynomial regressions or neural networks. An average of the first 12 producing months oil production at each well was chosen as the data to be modeled. PredictOnline (http://ford.nmt.edu/PredictOnline6/index. html) is an in-house devloped neural network package that is available online and is based on the fast-converging, feedforward, back-propagation conjugate gradient algorithm(Moller, 1993). A back-propagation feed-forward algorithm such as the conjugate gradient algorithm used here is “trained” using known inputs and outputs in an iterative fashion, with weights being sequentially adjusted until the desired tit (if possible) is achieved. There are a number of ways to determine which of a set of inputs (attributes) would best be used to form a regression for a particular output. Simply crossplotting each input against the output can give an indication of the quality of linear or multiple linear regression models that could be formed. For more complicated relationships found in many oil field problems such simple tools often do not provide adequate solutions. In previous studies we have used a single stage fuzzy-ranking algorithm to select inputs best suited for predicting the desired output(Balch et al., 1999; Balch et al., 2000; Hart et al., 2000; and Weiss, 2000-Z). The algorithm statistically determines how well a particular input could resolve a particular output with respect to any number of other inputs using fuzzy curve analysis. To illustrate the technique a simple example is given. Consider a set of random numbers in the range {O,l } using x={xi}, i=l 2 , ,. . .,99, and xi=O.Ol *i, and plot each value (yt= Random(xi)). Next add a simple trend to the random data (yi=(x$‘O.S+Random (xi)) and plot those values. For each data (xi, yi) a “fuzzy” membership function is defined using the following relationship: Fi(x) = exp( -(Xi b X))2)* yi Sample fuzzy membership functions are shown in Figures. 1 and 2. Here, b=O.l, since b is typically taken as about 10% of the length of the input interval of xi. A fuzzy curve was constructed using a summation of all individual fuzzy membership functions in (xi, yi), and this final curve can prioritize a set of inputs for linear or non-linear regressions. The fuzzy curve function is defined below: 4 R. S. BALCH, D. M. HART, W. W. WEISS, AND R. F. BROADHEAD 247 2F,(x)*~i FC(x) = i=’ ,~ C Fi.Cx> i=l Where N is the size of the data set or the total number of fuzzy membership functions. Figure 3 shows the curves for the data sets shown in Figs. 1 and 2. This simple example illustrates the ability of the fuzzy ranking approach to screen apparently random data for obscure trends such as the correlation between seismic attributes and reservoir properties4. More information is needed however to advance this analysis from the art of reading these fuzzy curves to a more robust and systematic elimination of less useful inputs, not only allowing selection of optimal inputs but also to allow an estimate of data quality and uniqueness. As such, we developed software based on a two-stage fuzzy ranking (Weiss, 2000-2). The twostage fuzzy ranking (TSFR) has two improvements: 1) Reduction of input variable space through random characterization and 2) setting hard rules for selection of best-input variables. TSFR introduces second stage fuzzy curves, with first and second stage fuzzy surfaces to select the most important and independent input variables for modeling, while removing the input variables that show random characteristics. TSFR uses first and second stage fuzzy curves to generate the fuzzy curve performance index (PC): With the addition of a known random variable into the input space the FCPI is normalized by the random Pc,a to produce the normalized fuzzy curve performance index (Pc,,v):
منابع مشابه
Using Artificial Intelligence to Predict Drilling Success Using Regional Data, Brushy Canyon Formation, Delaware Basin, New Mexico; #90023 (2003)
Incomplete and sparse information can introduce a high level of risk for oil exploration and development projects. “Expert” systems developed and used in several disciplines and industries have demonstrated beneficial results in modeling the decision making process of knowledgeable experts. A state-of-the-art exploration “expert” tool using a computerized data base and computer maps generated b...
متن کاملThe Fuzzy Expert Exploration Tool for the Delaware Basin: Development, testing and applications
The Delaware Fuzzy Expert Exploration Tool (FEE Tool) is an expert system designed to reduce exploration risk for the Lower Brushy Canyon formation of the Delaware Basin. The components of the Delaware FEE Tool include a knowledge base containing sets of rules developed through expert interviews, an answer base of numerical inputs to these rules, an inference engine that uses fuzzy logic to eva...
متن کاملHypogene Karst and Suflate Diagenesis of the Delaware Basin: Southeastern New Mexico and Far West Texas
v Page Introduction Geologic Setting Hydrologic Setting Hypogenic Speleogenesis Coffee Cave Discussion Conclusions Acknowledgements IV. CASTILE EVAPORITE KARST POTENTIAL MAP OF THE GYPSUM PLAIN, EDDY COUNTY, NEW MEXICO AND CULBERSON COUNTY, TEXAS: A GIS METHODOLOGICAL COMPARISON Abstract Introduction Geologic Setting Field Studies GIS Analyses Discussion Conclusions Acknowledgements V. EPIGENE ...
متن کاملReservoir characterization of a Permian deep-water sandstone, East Ford field, Delaware basin, Texas
Deep-water sandstones of the Delaware Mountain Group in west Texas and southeast New Mexico contained an estimated 1.8 billion bbl of original oil in place, but primary recovery from these fields is commonly less than 20%. East Ford field in Reeves County, Texas, which produces from the Ramsey sandstone in the upper Bell Canyon Formation, went directly from primary production to tertiary recove...
متن کاملApplication of an Adaptive Neuro-fuzzy Inference System and Mathematical Rate of Penetration Models to Predicting Drilling Rate
The rate of penetration (ROP) is one of the vital parameters which directly affects the drilling time and costs. There are various parameters that influence the drilling rate; they include weight on bit, rotational speed, mud weight, bit type, formation type, and bit hydraulic. Several approaches, including mathematical models and artificial intelligence have been proposed to predict the rate o...
متن کاملذخیره در منابع من
با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید
عنوان ژورنال:
دوره شماره
صفحات -
تاریخ انتشار 2003