Machine Learning Applied to SRV Modeling,Fracture Characterization,WellInterference and Production Forecasting in Low Permeability Reservoirs

摘要

The objective of this paper is to develop predictive models to optimize the(1)characterization of thestimulated reservoir volume(SRV),(2)discretization of the fracture network,and(3)hydraulic fracturingmodeling,by combining machine learning(ML)algorithms and reservoir engineering in low permeabilityreservoirs.An unsupervised learning algorithm is implemented to characterize the fracture network developed bymicro-seismic observations during hydraulic fracturing.A Self Organizing Map(SOM)and Multi-AttributeAnalysis are performed on the available seismic data to map the extension of the hydraulic fracturing stagesand the fracture network complexity in a low permeability reservoir.To correlate the mapped fracture network and discretized SRV,a 3D Finite Element Model(FEM)isdeveloped to estimate fracture behavior,stress response,and hydraulic fracture propagation,on the predictedand forecasted multi-attribute map of the reservoir.A 3D hydraulic fracture propagation model(HFPM)is introduced,to delimit the fracture geometry andremove data outliers in the SOM algorithm.Unsupervised algorithms rely on data quality.The efficiencyof hydraulic fracturing modeling is improved with a machine learning approach by refining the certaintyand quality of the data.An Artificial Neural Network(ANN)model helps to select the most significantparameters related to fracture modeling and simulation in the field.This approach allows us to recreate andforecast complex fracture networks in low permeability reservoirs,based on the learned geostatistical mapsand hydraulic fracturing parameters,particularly where the microseismicity is limited or unavailable.To validate the implementation of the 3D-HFPM in the field,an earthquake model is comparedwith statistically significant microseismic events obtained by the unsupervised iso-cluster algorithm.Therelationship showed a good agreement,which suggests the HFPM agrees with seismic observations in thefield.The machine learning application to fracture network modeling provides the capability to identifysusceptible areas to well interference and possible frac hits with higher certainty.This is so because theapproach improves the selection of seismic data and hydraulic fracturing parameters,employed to develop the complex fracture network in numerical commercial reservoir simulators.This helps to determinate thereservoir interconnectivity and flow patterns in the fracture network.This approach presents a robust manner for characterizing the SRV using a relative fast methodology,based on the combination of geostatistical and unsupervised learning modeling.The seismicity andhydraulic fracturing are connected using a multi-attribute and multi-disciplinary interpretation.It is apowerful tool for characterizing problematic fracture networks in unconventional reservoirs.