Description:

"Quantifying Uncertainty for Mapping Fracture Intensity: An Improved Workflow by Patrick Wong and Sean Boerner" Abstract: Fractured reservoir characterization requires understanding the spatial distribution of fracture intensity. Relationships between geological, structural, seismic attributes and fracture intensity are complex and highly nonlinear. This paper presents an improved integrated workflow for ranking attributes, understanding data limitations, mapping fracture intensity with neural networks, and quantifying uncertainty. The study uses soft attributes and cumulative gas production data as a proxy for fracture intensity at the Pinedale Anticline in Wyoming to demonstrate the workflow and its usefulness. Introduction: Modelling the spatial distribution of natural fractures in hydrocarbon-bearing reservoirs can be highly nonlinear. Fluid-flow behaviour is governed by fracture geometries (orientations, lengths, apertures), fracture intensity, and connectivity. For flow simulation, understanding dynamic relationships with the rock matrix (single dual porosity, single dual permeability) and applying a proper reservoir simulator are necessary for predicting flow performance. Fractured reservoir studies require deriving fracture intensity logs from static or dynamic data. Constraining property models with soft data improves predictions and lowers uncertainty in regions with poor well control. The resulting fracture intensity map can be used to constrain stochastic fracture planes via discrete fracture network (DFN) modeling, deriving various flow properties for each cell as inputs into a reservoir simulator. This paper focuses on mapping fracture intensity and quantifying the uncertainty of predictions. The improved workflow allows statistical ranking and selection of soft attributes, providing information on data set predictive power. A standard back-propagation neural network generates multiple stochastic maps of fracture intensity, then quantifies uncertainty relative to data limitations. The workflow is demonstrated with an application to the Pinedale Anticline in Wyoming, USA. Ranking Attributes: Regression-based mapping algorithms require a training set with known input-output pairs. For fracture intensity mapping, output is known fracture intensity from wells; inputs are soft attributes for mapping away from wells. Types of attributes vary by field (e.g., azimuthal AVO, stress-related, lithologies). Ranking attribute relevance allows selecting significant attributes for building regression models. Reducing input dimensions offers numerical advantages. In Wong and Boerner (2003b), a robust ranking algorithm called optimized piecewise rank correlation (OPRAC) was introduced. This algorithm uses rank correlation coefficients between attributes and output to rank significance, converting data values to ranks, constructing optimal linear segments through scatter-plots, calculating rank correlation coefficients for each segment, aggregating all coefficients into a weighted coefficient ranging from 0 to 1. Attribute Extrapolation: Regression-based methods integrate multi-attribute data but frequently encounter spatial and attribute extrapolations. Spatial extrapolation occurs when predictions are made away from regions with good well control; kriging variance quantifies uncertainty based on separation distances from hard data locations. Attribute extrapolation happens when full statistics of input attributes are not present in the training set. In Wong and Boerner (2003a), a coefficient of extrapolation was proposed to fine-tune binary extrapolation indicators. Individual attributes were examined, calculating relative distance from extreme values in training data; OPRAC coefficients weighed these distances, averaging for each cell with a value ranging from 0 to 1. Data within training set coverage had zero coefficient (no extrapolation). Extrapolation is now quantified as a degree. Mapping with Neural Networks: Neural networks have become important in geophysics (Wong et al., 2002; Aminzadeh and de Groot, 2004). They have been used successfully for fractured reservoir modelling (e.g., Zellou and Ouenes, 2001; Boerner et al., 2003) due to their abilities to learn, adapt, generalize from data. Multi-attribute integration and uncertainty quantification with multiple stochastic realizations offer practical advantages. Ключевые слова: fracture intensity, neural network, spatial distribution, coefficient, map, deviation, neural, natural fracture, quantifying, unsampled region, seismic attribute, regression model, type, geological, problem, work, fractured, extrapolation map, training data, anticline, article, point, relative distance, ranking, relevance, extent, content, reservoir characterization, model converged, hartley, oprac, correlation, data point, extreme, s-wave coherency, space, unsampled data, introduced, study, input, region, crack, crack density, extrapolation, field, mapping, reduction, wong boerner, result, sparse, regression-based method, break, workflow, cross-plot, break volume, training set, displayed, improved workflow, soft computing, high coefficient, distribution, poor, november, coefficient extrapolation, average, potential, unsampled, law, statistic, wyoming, spatial, demonstrate, fracture, std, location, method, avaz, network, output, curvature, conference, back-transformed, johnson, soft attribute, area, reservoir, modelling, correlation coefficient, technical, high, standard deviation, author, coverage, ability, standard, fractured reservoir, uncertainty, zellou, eage, seismic, pinedale anticline, technical article, stochastic, span, solution, proxy, attribute extrapolation, advantage, ha, hand, improved, limitation, s-wave, intensity, density, cell, attribute, data, volume, algorithm, degree, groot, boerner, wong, oprac coefficient, rank, highly, calculated, revealed, soft, model, quantifying uncertainty, coherency, input attribute, characterization, volume november, control, quantify, prediction, gauthier, set, lynn, training, number, paper, std map, wa, aminzadeh, pinedale, occurs, spe, mapping fracture, regression