Statistics

Key Takeaways

• Probabilistic reserve estimation using Monte Carlo simulation of the volumetric equation (GRV * N/G * phi * (1 - Sw) / Boi for oil) is the standard method for quantifying the uncertainty in recoverable reserves from exploration and early appraisal wells, where the reservoir parameters are uncertain and must be described by probability distributions rather than single deterministic values: each input parameter — gross rock volume (from seismic interpretation of the closure area), net-to-gross ratio (from analogy wells or forward seismic modeling), porosity (from log analysis with uncertainty bounds), water saturation (from resistivity interpretation with Archie equation parameter uncertainty), and recovery factor (from analogy fields or reservoir simulation) — is assigned a probability distribution (log-normal, normal, triangular, or uniform depending on the parameter and available data) that reflects the current state of knowledge about that parameter; Monte Carlo simulation randomly samples each input distribution many thousands of times and computes the resulting recoverable volume for each sample, producing a probability distribution of recoverable volumes that is characterized by the P10 (the volume exceeded by 90% of simulations, a high estimate), P50 (the median, exceeded by 50%), and P90 (a low estimate exceeded by only 10%); the ratio of P10 to P90 is the uncertainty range, which is typically 3:1 to 20:1 for exploration prospects where the subsurface is poorly constrained, and narrows to 1.5:1 to 3:1 for developed fields where extensive well data has reduced the key uncertainties; SEC (Securities and Exchange Commission) reserve classification rules require that proved reserves be estimated with high certainty (roughly the P90 level), while probable and possible reserves correspond to P50 and P10 estimates.
• Decline curve analysis applies nonlinear regression to production rate-time data to fit the Arps empirical decline equation (q = qi / (1 + b*Di*t)^(1/b), where q is the current rate, qi is the initial rate, Di is the initial decline rate, b is the Arps decline exponent, and t is the elapsed time) and to extrapolate the fitted curve to forecast future production and estimate ultimate recovery: the Arps equation has three limiting cases that are distinguished by the value of b — exponential decline (b = 0, constant fractional decline rate, typical of volumetric reservoirs and fully depletion-drive systems), hyperbolic decline (0 < b < 1, declining fractional decline rate, typical of transient or mixed-drive systems), and harmonic decline (b = 1, the limiting case of hyperbolic with the slowest possible rate of decline decrease); the choice of decline type and the values of qi, Di, and b are estimated by fitting the Arps equation to the historical production data using nonlinear least-squares regression, and the statistical goodness of fit (R-squared, residual plots, confidence intervals on the fitted parameters) determines whether the fitted parameters are reliable for long-range extrapolation; in unconventional tight oil and gas wells, the early transient flow period (months to years) produces hyperbolic decline with b values of 1.5-2.5 (super-Arps behavior that violates the theoretical maximum of b = 1 for pressure depletion systems) that overestimates long-term production if extrapolated without terminal exponential decline correction; the modified hyperbolic-to-exponential transition model (fitting hyperbolic decline in the early period and transitioning to exponential decline when the decline rate falls to 5-10% per year) is the industry standard for unconventional well EUR estimation.
• Bayesian statistics provides the formal mathematical framework for incorporating prior knowledge (from analogy fields, geological models, or expert judgment) with new data (from drilling results, production history, or seismic reprocessing) to update the probability distributions of uncertain parameters in a way that is mathematically consistent with the rules of probability: Bayes' theorem states that the posterior probability distribution p(theta | data) is proportional to the product of the likelihood p(data | theta) and the prior probability p(theta), where theta represents the uncertain parameter(s) and data represents the new observations; in exploration, the prior probability of success (PoS) for an undrilled prospect is estimated from geological risk assessment (trap integrity, seal quality, source rock quality, reservoir quality, migration pathway), and the posterior PoS is updated as new data becomes available (nearby well results, new seismic attributes, production data from wells in the same play); the Bayesian framework explicitly accounts for the difference between the information that was available at the time of the prior estimate and the information provided by the new data, preventing the cognitive bias of anchoring to the prior estimate when new evidence contradicts it; petroleum companies that apply Bayesian risk assessment to their exploration portfolios can calibrate their geological models by comparing predicted PoS values to actual drilling success rates (the fraction of drilled prospects that discovered commercial hydrocarbons), and use the calibration results to identify systematic biases in their risk assessment methodology (such as consistent overestimation of trap integrity or seal quality in a particular play type).
• Geostatistics extends classical statistics to spatial data by incorporating the spatial correlation structure of reservoir properties (the variogram, which measures how the similarity between property values at two locations decreases as the distance between those locations increases) into the interpolation and simulation of properties between wells: kriging (ordinary, simple, and indicator kriging) is the geostatistical interpolation method that provides the minimum variance unbiased linear estimator of a property at an unsampled location, using the variogram to determine the optimal weights assigned to the surrounding sample values; sequential Gaussian simulation (SGS) is the geostatistical stochastic simulation method that generates multiple equally probable realizations of the reservoir property distribution, each conditioned to the well data and consistent with the variogram model, providing an ensemble of model realizations that together quantify the uncertainty in the spatial distribution of reservoir properties beyond the wells; the choice of variogram range (the distance beyond which values become spatially uncorrelated) and nugget effect (the proportion of variance that is spatially uncorrelated at very short separation distances) has a major impact on the connectivity of high-permeability sand bodies in simulation realizations and therefore on the simulated recovery factor and breakthrough timing; incorrect variogram parameters (too short a range, too high a nugget) produce simulation realizations with unrealistically disconnected sand bodies that underestimate sweep efficiency and overestimate production risk.
• Statistical process control (SPC) applied to drilling and production operations monitors whether a process is operating within its expected performance envelope and detects anomalies that indicate equipment failure, operational deviation, or formation change before they develop into significant problems: control charts (Shewhart charts, CUSUM charts, EWMA charts) plot measured process variables (drilling rate of penetration, mud weight, pump pressure, gas reading, production rate, water cut) versus time and alert the operator when values exceed control limits (typically mean plus or minus three standard deviations for a Shewhart chart) or when a series of sequential values shows a non-random trend (seven consecutive points above the mean, indicating a systematic shift in the process); in real-time drilling monitoring centers (remote operations centers, ROC), statistical alarms on downhole measurement-while-drilling (MWD) data (pressure, temperature, vibration, inclination) alert the drilling engineer to conditions such as packoff (sudden increase in annular pressure above the expected value), motor stall (drop in differential pressure across the mud motor), or bit bounce (high-amplitude vibration in a characteristic frequency band) that require immediate action to prevent lost-time incidents; the statistical threshold values for these alarms are calibrated from historical drilling data for similar formations and well designs, with false alarm rates (alarms triggered by normal process variability rather than genuine anomalies) balanced against missed detection rates (genuine anomalies not flagged by the alarm system) using receiver operating characteristic (ROC) curve analysis.