MACHINE LEARNING FOR PRODUCTION FORECASTING: ACCURACY THROUGH UNCERTAINTY 7 TH RESERVES ESTIMATION UNCONVENTIONALS JUNE 20 22, 2017 HOUSTON, TX DAVID FULFORD APACHE CORPORATION
PRODUCTION FORECASTING IN UNCONVENTIONALS 2
WE NOTICED A PROBLEM 3
IDENTIFYING CAUSES SOLUTIONS Machine Learning forecasting becomes standard process
OUTLINE Problem Statement Key Components Model Regression / Parameter Estimation Machine Learning Representative Forecasts & Type Wells Workflow 5
PROBLEM STATEMENT Our forecasting was yielding unreliable results Low permeability leads to long duration transient regimes Production decline behavior differs significantly in transient vs. transitionary vs. boundary-dominated regimes Most empirical models are formulated to forecast a single, specific regime 6
KEY COMPONENTS What should production forecasts for unconventionals look like? Develop a new model Transition away from best fit to most-likely fit Requires prior belief of production forecast behavior Conduct look-backs, calibration of prior beliefs, verification of forecasts Automation & Workflow Improvement
CHALLENGES IN CHANGE MANAGEMENT Apache Reservoir Engineers had to learn a new workflow to develop skill in estimation: Understand what s meant by forecasting bias
AN ASIDE ON BIAS What is bias? The difference between a estimator s expected value and the true value of the parameter being estimated Why be explicit about bias? It s necessary!!! It s quantified and therefore can be: Measured Auditable Tracked Repeatable Calibrated Consistent
CHALLENGES IN CHANGE MANAGEMENT Apache Reservoir Engineers had to learn a new workflow to develop skill in estimation: Understand what s meant by forecasting bias Practice estimation of bias with regards to unobserved production data as opposed to best-fit of observed data Develop discipline in self-calibration Estimate Evaluate Calibrate Repeat!!!
DEVELOPING A IMPROVED MODEL We don t know specifically how multi-fractured horizontal wells (MFHW) will behave in the long-term But we can bound the problem! Successful attempts are commonly applied as part of a model-based rate-transient analysis (RTA) process. However, requires much more data than rate-time, and much more time not realistic for a reserves estimation workflow How do we combine the physics of these models with the time and data limitations imposed upon the reserves estimation workflow?
DEVELOPING A IMPROVED MODEL Analytic Model Planar Fractures Trilinear planar fracture Brown et al. (2011)
DEVELOPING A IMPROVED MODEL Analytic Model Complex Fractures Enhanced Fracture Region (EFR) induced fracture natural fracture proppant placement Stalgorova and Mattar (2012)
SUMMARY OF ANALYTIC MODEL BEHAVIOR Fracture Linear How does a single parameter (b-parameter) forecast all of this? Transitionary Regime Enhanced Permeability Effect (Linear) No Fracture Linear Enhanced Permeability Effect (Linear) Transitionary Regime Transitionary Regime Inner Matrix Linear Outer Matrix Linear Inner Matrix Linear Outer Matrix Linear Matrix Boundary Dominated Matrix Boundary Dominated Fulford (2017)
EMPIRICAL APPROXIMATION OF BEHAVIOR b i b f b terminal Transient Linear Flow Inter-fracture pressure interference Boundary Influenced Flow Compound Linear Flow Boundary Dominated Flow b = 2 b < 1 b = 2 b < 1 Pressure Depletion in SRV Inter-well pressure interference Pressure Depletion in Reservoir 15 After Song & Ehlig-Economides (2011)
EMPIRICAL APPROXIMATION OF BEHAVIOR b i 2.5x b f 2.5x b terminal Complex physics captured by assumptions imbedded in model: No fracture-linear flow Finite FCD b i 2 b terminal = 0.3 if oil 0.5 if gas t terminal = 5 yrs 10 yrs Fit parameters: b f t elf time to end of linear flow time to end of b i
EMPIRICAL VALIDATION Eagle Ford Marmaton Granite Wash Woodford 17
EMPIRICAL VALIDATION Wolfcamp Cleveland Bakken Bluesky Kaybob 18
KEY COMPONENTS What should production forecasts for unconventionals look like? Transient Hyperbolic Model Not derived from, but follows from, first principles Transition away from best fit to most-likely fit Estimate bias w.r.t. to unobserved production data Conduct look-backs, calibration of prior beliefs, verification of forecasts Quantified bias + machine learning 19
MACHINE LEARNING FOR PRODUCTION FORECASTING 20
WHAT IS MACHINE LEARNING? Machine learning is a name given to algorithms and techniques for the extraction of predictive models from data Unsupervised learning extracts structure, grouping, and dimension reduction from correlations and clusterings in unlabeled data Supervised learning works in labeled data sets to learn models which can predict values for future unobserved data map one or more predictors to one or more targets 21
NOT AUTOMATED FORECASTING Our problem: given observed data of well performance vs. time, learn an appropriate forecast model to predict future well performance Time-rate data are indirect observations of fluid & rock properties Non-unique inverse problem Many local optimums such that a least error fit is not a best fit and does not yield a best forecast 22
SOLUTION SPACE OF INVERSE PROBLEMS Markov Chain Monte Carlo Simulation Proven technology with 20+ years of use in oilfield Reservoir simulation Model selection Model calibration Uncertainty quantification Seismic inversion / seismic processing Estimation of model parameters required to guide/constrain the solution set 23
MARKOV CHAIN MONTE CARLO Other applications Cryptography Text decryption Astronomy Analysis of CMB to determine age of the Universe Meteorology Hurricane risk assessment Chemistry Nanoscale research in phase behavior etc. etc. etc. 24
INFERENCE MCMC Inference is inherently Bayesian Bayes Theorem infers between our prior knowledge & beliefs and new data we acquire. Avoids base rate fallacy disregarding general information and overvaluing specific information THM provides a generalized solution for time-rate performance Production history provides specific data on time-rate performance From belief of the general behavior, converge to the specific behavior 25
Rate 7 th Reserves Estimation Unconventionals June 20 22, 2017 Houston, TX DISTRIBUTION OF FORECASTS 10,000+ possible forecasts are summarized into discrete percentiles Actual vs. MCMC Forecasts 26 Time
Rate 7 th Reserves Estimation Unconventionals June 20 22, 2017 Houston, TX DISTRIBUTION OF FORECASTS Possible fits of data + uncertainty of future performance Actual vs. MCMC Forecasts 27 Time
DISTRIBUTION OF FORECASTS Representative Forecasts What are the features of the set of forecasts that recover the P50 volume? Log Rate vs. Log Time Log Rate vs. Time 28
RESERVE REVISIONS How much should I expect to revise forecasts from month-to-month with this approach? On average, zero! Change in EUR from prior month Clifford and Torres (2017)
TYPE WELL CREATION Analytic creation of Type Wells provides statistical validity Free lunch Any set of forecasted wells can be used to create a type well bias is inherited from individual well forecasts Clifford and Torres (2017)
TYPE WELL VALIDATION Type Well by Vintage Clifford and Torres (2017)
WORKFLOW IMPROVEMENT Moves empirical forecasting toward the scientific method Bias estimation = hypothesis testing Focus shifts from curve fitting to interpretation Diagnostics improve forecast accuracy -> forecast confidence Wells are samples from a population Behavior of one well provides evidence for behavior of other wells Type wells are the expected value of the sampled set of wells
WORKFLOW TIME INVESTMENT 7 th Reserves Estimation Unconventionals June 20 22, 2017 Houston, TX
APACHE IMPLEMENTATION Two primary implementations Excel-based utility for daily data Spotfire-based utility for public data Production data acquisition automated via SQL query Batch processing of wells e.g. forecast entire field in a few minutes or basin in < 1 hour Automated export to ARIES for all phases 34
BIG DATA ANALYTICS 7 th Reserves Estimation Unconventionals June 20 22, 2017 Houston, TX
DISCUSSION & CONCLUSIONS Machine Forecasting outperforms Human Forecasting More rigorous model & quantified bias no big surprise! Not because it replaces humans but enhances engineering judgment Workflow doesn t replace ARIES (or PEEP or Val. Nav., etc.) Seamless integration into pre-existing Apache processes
Questions? 37
REFERENCES Brown, M., Ozkan, E., Raghavan, R., and Kazemi, H. 2011. Practical Solutions for Pressure-Transient Responses of Fractured Horizontal Wells in Unconventional Shale Reservoirs. SPE Res Eval & Eng 14 (6): 663 676. SPE-125043-PA. http://doi.org/10.2118/125043-pa. Clifford, S., and Torres, T. 2017. Using a Systematic, Bayesian Approach to Unlock the True Value of Public Data; Midland Basin Study. Presented at Unconventional Resources Technology Conference in Austin, TX, USA, 24 26 July. URTeC-2697318. http://doi.org/10.15530/urtec-2017-2697318. Fulford, D.S., and Blasingame, T.A. 2013. Evaluation of Time-Rate Performance of Shale Wells Using the Transient Hyperbolic Relation. Presented at SPE Unconventional Resources Conference in Calgary, Alberta, Canada, 5 7 November. SPE-167242- MS. http://doi.org/10.2118/167242-ms. Fulford, D.S., Bowie, B., Berry, M.E., and Bowen, B. 2016. Machine Learning as a Reliable Technology for Evaluating Time/Rate Performance of Unconventional Wells. SPE Econ & Mgmt 8 (1): 23 29. SPE-174784-PA. http:/doi.org/10.2118/174784-pa. Song. B, and Ehlig-Economides, C.A., 2011. Rate-Normalized Pressure Analysis for Determination of Shale Gas Well Performance. Presented at SPE North American Unconventional Gas Conference and Exhibition in The Woodlands, Texas, USA, 14 16 June. SPE-144031-MS. http://doi.org/10.2118/144031-ms. Stalgorova, E., and Mattar, L. 2012. Pratical Analytical Model to Simulate Production of Horizontal Wells with Branch Fractures. Presented at SPE Canadian Unconventional Resources Conference in Calgary, Alberta, Canada, 30 October 1 November. SPE-162515-MS. http://doi.org/10.2118/162516-pa. 38
APPENDIX 39
LINEAR FLOW DURATION 40 φ = 4% k =.0001 md
DISTRIBUTION OF FORECASTS Posterior PDF Posterior CDF 30-yr Cumulative Production, Mbbl 41
DISTRIBUTION OF FORECASTS Reasonable assumption of linearity of parameters to EUR Parameter Linearity 30-yr Cumulative Production Residual of Linear Fits 30-yr Cumulative Production 42
EMPIRICAL APPROXIMATION OF BEHAVIOR Enhanced Permeability Effect (Linear) b i b f b terminal Outer Matrix Linear Inner Matrix Linear Matrix Boundary Dominated