Geir Evensen History matching reservoir models using the EnKF/EnKS Hydro: Aluminium and energy supplier Founded 1905 36 000 employees in 40 countries Leading offshore oil producer Worlds third largest aluminium supplyer Pioneer in renewable energy and energy efficient solutions Date: 2004-01-16 • Page: 2 Hydro Research Centre, Bergen Programs Improved 60% Exploration Recovery Date: 2004-01-16 • Page: 3 Small Fields Polar Regions Climate & Renewable energy Program: 60% Recovery Extend lifetime of mature fields mm cm m Projects: Seismic mapping Reservoir geophysics IOR Oseberg seismics Reservoir characterization km Methods for tertiary recovery Well and intervention technology Unstable flow Date: 2004-01-16 • Page: 4 Introduction Simulation model must be consistent with the reservoir production history to be used for predicting future production. History matching = Tuning of model parameters or parameter estimation. For history mathcing we must: Define the parameter space (porosity, permeability, reservoir structure, transmissibilities,….). Define a cost fuction (prior statistical weights, norm, etc.). Chose a minimization method (gradient descent, Monte Carlo, EnKF). Date: 2004-01-16 • Page: 5 Minimization problem Highly nonlinear, multiple local minima. Hard to solve using traditional minimization methods. Data errors independent in time and model a Markov sequence: Reformulation as a sequence of minimization problems. Each subproblem provides a prior for the next problem. Equivalent to solving the full problem in one go for linear models. Evensen 2005 (www.nersc.no/~geir) Approach used by the EnKF/EnKS. Date: 2004-01-16 • Page: 6 Update procedure in EnKS/EnKF d_1 d_2 Prediction EnKF update EnKS update Updates Date: 2004-01-16 • Page: 7 d_3 Time Nonlinear minimization problem Date: 2004-01-16 • Page: 8 Ensemble Kalman Filter (EnKF) Statistical method which solves the combined state and parameter estimation problem. Monte Carlo version of the original Kalman Filter. EnKF home page: http://www.nersc.no/~geir/EnKF The EnKF is now used for ocean and weather prediction TOPAZ system: http://topaz.nersc.no Meteorological Service of Canada Potential for history matching and reservoir management? Date: 2004-01-16 • Page: 9 Ensemble Kalman Filter (EnKF) Representation of uncertainty using an ensemble of model solutions and randomized measurements: Uncertainty predictions by ensemble integration: Describes error covariances for the model estimate and measurements. Each ensemble member is integrated using a stochastic version of the dynamical model. Includes the impact of model errors. Model ensemble is updated from measurements using a variance minimizing analysis scheme: Assumption about Gaussian error statistics for prediction. Date: 2004-01-16 • Page: 10 Measurements with errors True state Initial state with errors Time Date: 2004-01-16 • Page: 11 Measurements with errors True state Model prediction with errors Time Date: 2004-01-16 • Page: 12 Measurements with errors True state Updated estimate with errors Model prediction with errors Time Date: 2004-01-16 • Page: 13 Measurements with errors True state New model prediction with errors Updated estimate with errors Time Date: 2004-01-16 • Page: 14 Measurements with errors True state Updated estimate with errors Time Date: 2004-01-16 • Page: 15 EnKF-Eclipse Implementation at Hydro Suite of F90 programs controlled by unix script Simulation software for generation of the initial ensemble. Only used to integrate model states forward in time. interface between ECLIPSE and the EnKF: Simulation of realizations based on statistical geomodel. Conditioning on log and AI data. Eclipse simulator is used as a “black box”: Modular and model independent implementation. Based on generic EnKF distribution from http://www.nersc.no/~geir/EnKF/ Convert between ECLIPSE restart files and ensemble files. Differentiate between dynamic and static variables. EnKF analysis program. Date: 2004-01-16 • Page: 16 EnKF-Eclipse Implementation: Initialization of ensemble Eclipse data file eclipse Statistical geomodel iniens logobs.uf Date: 2004-01-16 • Page: 17 Grid and restart info Poro/perm ensemble EnKF ens0F.uf ens0S.uf ens0A.uf ens0S.uf EnKF-Eclipse Implementation Ensemble integration and update For j=0, data_times For i=1,nrens ens{j}A.uf ens{j}S.uf Eclipse restart file interface interface Eclipse restart file ens{j+1}F.uf ens{j+1}S.uf done ens{j+1}F.uf EnKF obs{j+1}.uf done Date: 2004-01-16 • Page: 18 ens{j+1}A.uf eclipse Estimation of permeability using EnKF Reference First guess Major structure of permeability field is recovered! Date: 2004-01-16 • Page: 19 Estimate 25 months Estimation of permeability using EnKF Permeability Information from data accumulates in time. Result is consistent with statistical input model. Definition of initial statistical model is crucial. Cross section distance Results from idealized experiments are very promising! Date: 2004-01-16 • Page: 20 Some previous EnKF reservoir applications Nævdal, Mannseth and Vefring (2003), Near well reservoir monitoring through EnKF, (SPE 75235). Nævdal, Johnsen, Aanonsen and Vefring (2003), Reservoir monitoring and continuous model updating using EnKF, (SPE 84372). Gu and Oliver (2004), History matching of the PUNQ-S3 reservoir model using the EnKF, (SPE 89942). Gao, Zafari and Reynolds (2003), Quantifying uncertainty for the PUNQS3 problem in a Bayesian setting with RML and EnKF, (SPE 93324). Liu and Oliver (2005), Critical evaluation of the EnKF on history matching of geological facies, (SPE 92867). Wen and Chen (2005), Real-time reservoir model updating using EnKF, (SPE 92991). All conclude positively regarding the use of EnKF for history matching. All consider idealized experiments. Date: 2004-01-16 • Page: 21 EnKF-Eclipse for Oseberg Alpha North Complex reservoir. 5 producers. 3 injectors. Simulation model hard to match. Uncertainty about faults, barriers, channels, and communication with other reservoirs. Date: 2004-01-16 • Page: 22 Date: 2004-01-16 • Page: 23 Some preliminary remarks Good match for GOR at wells started early in simulation. Poor match for wells started late in simulation. Easy to fit unconstrained model to the first production data. Too large weight on early production data? Neglected model errors? Typical parameterization problem? Possible to include faults, structural errors, transmissibilities? Uncertainty in production data. 10-20% or more? QC of production data crucial. Date: 2004-01-16 • Page: 24 Summary Ongoing field case with new Oseberg alpha north model: Includes pluri-gaussian facies representation. Assimilates production and 4D seismic data. Ongoing field case for Rimfaks: Collaboration with Statoil. Date: 2004-01-16 • Page: 25