Multivariate Pore-Pressure Prediction And Uncertainty Analysis

More hydrocarbon exploration and development are conducted in over-pressured zones as offshore hydrocarbon exploration gradually moves to deep sea.Predrill abnormal pore-pressure prediction has become a key step in the hydrocarbon exploration and development.Precise prediction of abnormal pore-pressure is of great importance in hydrocarbon exploration,oil field development and reservoir engineering.The most used methods of pore-pressure prediction are those based on univariate empirical relation between geophysical property and pore pressure(or effective stress).Abnormal pore pressure has complex generation mechanisms,and is affected by regional structural background,depositional characteristics,and other factors.Univariate model often cannot adequately describe the complex changes of pore pressure.This dissertation conducts research on methods of multivariate pore-pressure prediction.It aims at building a machine learning method for multivariate pore-pressure prediction with well log data and a machine learning method for multivariate pore-pressure prediction with seismically generated petrophysical properties from pre-stack AVO inversion and reservoir property inversion.Moreover,seismic pore-pressure prediction has considerable uncertainty.Quantitative analysis of the uncertainty of seismic pore-pressure prediction can help reduce the risk of drilling.The theory and methods of Geostatistics provide a methodology for quantitative analysis of uncertainty of geologic processes.Hence,this dissertation conducts research on quantitative analysis of seismic pore-pressure prediction with geostatistical simulation.The main research results of the dissertation include:1)With regressional analysis of experimental data of rock samples on the relation between velocity and effective stress,Eberhart-Phillips' model with the functional form of the combination of linear and exponential describes the relation for sand and Sayers' model with the functional form of power-law describes better the relation for shale.The functional form of the two models are strong assumptions of lithology.The use of nonparametric model can release the restriction of functional forms,hence assumptions of lithology,and describes the relation of petrophysical properties and effective stress more precisely.2)Base on the nonparametric multivariate effective stress model,a machine learning method for pore-pressure prediction with well log data is proposed.Applying machine learning models require a large number of training examples,measured pressures which is often few in number cannot provide enough training examples.The use of effective stress data in the normally compacted intervals instead of measured pressures for building the training dataset solve the problem of inadequate data size.Since the distribution of values of porosity and shale volume is skewed in nature,the quantile scaler is a more effective method for normalization of the training dataset than the Min/Max scaler.The analysis on the results of hyperparameter optimization of different machine learning models shows that the same set of optimized hyperparameters can be shared across wells with the same geological background.Compared with multilayer perceptron neuron network,support vector machine and gradient boosting,random forests perform better in pore-pressure prediction with well log data.The performance of machine learning models on pore-pressure prediction correlates with goodness of fit and generalization ability.Compared with parametric methods,the machine learning method give more precise top of geopressure.3)A machine learning method for pore-pressure prediction with seismic data is proposed.It uses prestack simultaneous inversion to obtain velocity for pore-pressure prediction and density for overburden pressure calculation,and uses petrophysical inversion to obtain porosity and shale volume for pore-pressure prediction.Then the machine learning method for pore-pressure prediction is applied on each CDP to construct the pore-pressure field.Since pore-pressure prediction with seismic data is a computation intensive task,using gradient boosting instead random forest can drastically reduce computation time with little loss of precision.When correction for unloading pressure,since unloading exponent U can only be obtained on well data,kriging is used to construct the regional spread of unloading exponent.The analysis of predicted pressure and measured pressure shows that the predicted pore-pressure by the proposed method shows good precision and comply with geologic background.4)Since seismic velocity,porosity and shale volume are correlated variables,cosimulation is needed reconstruct their spatial correlation.However,co-simulation of several variables is complex and inefficient.This dissertation proposes to use cosimulation with MAF transforms to co-simulate and obtain many realizations of velocity,porosity and shale volume.Hence the machine learning method for pore-pressure prediction with seismic data is used to obtain many realizations of predicted pore pressure.Then the local distribution of predicted pore pressure can be determined.With the proposed method,a confidence interval can be obtained to quantify the uncertainty.The innovation points of the dissertation include:1)A machine learning method for multivariate pore-pressure prediction with well log data is proposed.It is based on the nonparametric transform of effective stress and uses machine learning to predict pore pressure with well log data.2)A machine learning method for multivariate pore-pressure prediction with seismic data is proposed.With pre-stack simultaneous inversion and petrophysical inversion,it applies machine learning method to pore-pressure prediction with seismic data.3)A method for quantifying the uncertainty of multivariate pore-pressure prediction with geostatistical simulation is proposed.It uses MAF transform and machine learning method for pore-pressure prediction with seismic data to obtain realizations of predicted pore pressure,then use the local distribution of predicted pore pressure to quantify uncertainty.