Uncertainty reduction in reservoir parameters prediction from multiscale data using machine learning in deep offshore reservoirs.

Arigbe, Ovoke Daniel

Abstract

Developing a complete characterization of reservoir properties involved in subsurface multiphase flow is a very challenging task. In most cases, these properties - such as porosity, water saturation, permeability (and their variants), pressure, wettability, bulk modulus, Young modulus, shear modulus, fracture gradient - cannot be directly measured and, if measured, are available only at small number of well locations. The limited data are then combined with geological interpretation to generate a model. Also increasing the degree of this uncertainty is the fact that the reservoir properties from different data sources - like well logs, cores and well test - often produce different results, thus making predictions less accurate. The present study focussed on three reservoir parameters: porosity, fluid saturation and permeability. These were selected based on literature and sensitivity analysis, using Monte Carlo simulations on net present value, reserve estimates and pressure transients. Sandstone assets from the North Sea were used to establish the technique for uncertainty reduction, using machine learning as well as empirical models after data digitization and cleaning. These models were built (trained) with observed data using other variables as inputs, after which they were tested by then using the input variables (not used for the training) to predict their corresponding observed data. Root Mean Squared Error (RMSE) of the predicted and the actual observed data was calculated. Model tuning was done in order to optimize its key parameters to reduce RMSE. Appropriate log, core and test depth matching was also ensured including upscaling combined with Lorenz plot to identify the dominant flow interval. Nomographic approach involving a numerial simulation run iteratively on multiple non-linear regression model obtained from the dataset was also run. Sandstone reservoirs from the North Sea not used for developing the models were then used to validate the different techniques developed earlier. Based on the above, the degree of uncertainty associated with porosity, permeability and fluid saturation usage was demonstrated and reduced. For example, improved accuracies of 1-74%, 4-77% and 40% were achieved for Raymer, Wyllie and Modified Schlumberger, respectively. Raymer and Wyllie were also not suitable for unconsolidated sandstones while machine learning models were the most accurate. Evaluation of logs, core and test from several wells showed permeability to be different across the board, which also highlights the uncertainty in their interpretation. The gap between log, core and test was also closed using machine learning and nomographic methods. The machine learning model was then coded into a dashboard containing the inputs for its training. Their relationship provides the benchmark to calibrate one against the other, and also to create the platform for real-time reservoir properties prediction. The technology was applied to an independent dataset from the Central North Sea deep offshore sandstone reservoir for the validation of these models, with minimum tuning and thus effective for real-time reservoir and production management. While uncertainties in measurements are crucial, the focus of this work was on the intermediate models to get better final geological models, since the measured data were from the industry.

Citation

ARIGBE, O.D. 2020. Uncertainty reduction in reservoir parameters prediction from multiscale data using machine learning in deep offshore reservoirs. Robert Gordon University [online], PhD thesis. Available from: https://openair.rgu.ac.uk

• Thesis Type: Thesis
• Deposit Date: Jul 16, 2020
• Publicly Available Date: Jul 16, 2020
• Keywords: Oil and gas reservoirs; Reservoir properties; Risk management; Sandstone; Operational analytics
• Public URL: https://rgu-repository.worktribe.com/output/947866
• Award Date: Jun 30, 2020