In complex carbonate reservoirs, it is crucial to understand the connections between reservoir compositions (minerals, facies, and properties). Conventionally, core samples have been used to measure reservoir parameters and identify minerals and facies. However, core samples are limited to certain wells. Therefore, additional techniques are necessary to overcome this limitation comprehensively. This study aims to identify key mineralogical and facies components of the Barra Velha Formation (BVR) and their relation to reservoir parameters. Dolomite, calcite, quartz, and clay minerals were commonly found using X-ray Diffraction (XRD). By employing multi-mineral (MM) petrophysical evaluations, we accurately recreated mineral quantities from XRD and petrophysical properties from core analysis to ensure reliability. Replications of inputs well logs and the mineralogical volume from spectroscopic (ECS) were used as reliability techniques for validating the MM. A total of 47 wells were analyzed using those methods. In this study, the classification of facies was accomplished through the selection of three prominent supervised artificial intelligence techniques, among which SOM, a widely employed method for facies estimation, was included. Additionally, the ensemble methods of Random Forest and XGBoost were adopted due to their recognized efficacy in handling tabular data and their track record of success in machine learning and artificial intelligence competitions. Remarkably, the performance evaluation revealed that Random Forest and XGBoost algorithms outperformed SOM, yielding the most favorable outcomes in this context. An integrated analysis of mineralogical and facies results was conducted, incorporating production data and special profiles such as nuclear magnetic resonance (NMR) and Wellbore Image (WBI) to identify vug-containing areas. The dolomitic facies exhibited favorable reservoir qualities, influenced by diagenetic processes represented by vuggy porosity, which enhanced permeability. Shrubstones, spherulites, and reworked facies showed superior petrophysical qualities and were connected with productive regions, leading to elevated dolomite concentrations, and vuggy abundance. The study highlights two major innovations: the use of mineralogical volume from multi-mineral assessments as inputs for AI-based property estimation to improve facies estimates, and the discovery of relationships between facies, minerals, and reservoir properties, compared to production data. This understanding allows for more accurate static model creation, optimal production interval selection, improved hydrocarbon recovery, and better specification of stimulation processes.
Tag: Machine Learning
Effects of the Random Forests Hyper-Parameters in Surrogate Models for Multi-Objective Combinatorial Optimization: A Case Study using MOEA/D-RFTS
Surrogate models are techniques to approximate the objective functions of expensive optimization problems. Recently, Random Forests have been studied as a surrogate model technique for combinatorial optimization problems. Nonetheless, Random Forests contain several hyper-parameters that are used to control the prediction process. Despite their importance, research on the effects of these hyper-parameters is scarce. Therefore, this paper performs a systematic investigation of the effects of different combinations of values for the Random Forest hyper-parameters on the approximation of well-known multi-objective combinatorial benchmark problems. The results show that the number of samples to consider when building each tree and the minimum number of samples to be at the leaf node are the two most important hyper-parameters in this context.
Data-driven forecasting of oil & gas production: a recurrent neural networks approach
Our main objective is to develop and compare two data-driven forecasting models based on Recurrent Neural Networks (RNN). We built the models with the Long Short-Term Memory (LSTM) and the Gated Recurrent Units (GRU). The data was obtained from a synthetic benchmark (UNISIM-II-H), which simulates an Oil & Gas (O&G) pre-salt production field. We performed the training/testing procedures: we trained the models with 80% of data from the same well and tested with the remaining data. Forecasting is calculated using as input the historical record of production variables (liquid, oil, gas & pressure). We also measured the symmetric mean absolute percentage error (SMAPE) to compare our forecasting with the data from the selected benchmark. Several experiments were performed; we used 28, 45, and 90 days for historical records, with 7, 15, and 30 days for forecasting, respectively—most of the experiments exhibited and SMAPE lower to 20. Results from the RNN models exhibited relative values compared to the expected data from the benchmark in most experiments for oil & gas production values.
Selection of Representative Scenarios Using Multiple Simulation Outputs for Robust Well Placement Optimization in Greenfields
In greenfield projects, robust well placement optimization under different scenarios of uncertainty technically requires hundreds to thousands of evaluations to be processed by a flow simulator. However, the simulation process for so many evaluations can be computationally expensive. Hence, simulation runs are generally applied over a small subset of scenarios called representative scenarios (RS) approximately showing the statistical features of the full ensemble. In this work, we evaluated two workflows for robust well placement optimization using the selection of (1) representative geostatistical realizations (RGR) under geological uncertainties (Workflow A), and (2) representative (simulation) models (RM) under the combination of geological and reservoir (dynamic) uncertainties (Workflow B). In both workflows, an existing RS selection technique was used by measuring the mismatches between the cumulative distribution
of multiple simulation outputs from the subset and the full ensemble. We applied the Iterative Discretized Latin Hypercube (IDLHC) to optimize the well placements using the RS sets selected from each workflow and maximizing the expected monetary value (EMV) as the objective function. We evaluated the workflows in terms of (1) representativeness of the RS in different production strategies, (2) quality of the defined robust strategies, and (3) computational costs. To obtain and validate the results, we employed the synthetic UNISIM-II-D-BO benchmark case with uncertain variables and the reference fine- grid model, UNISIM-II-R, which works as a real case. This work investigated the overall impacts of the robust well placement optimization workflows considering uncertain scenarios and application on the reference model. Additionally, we highlighted and evaluated the importance of geological and dynamic uncertainties in the RS selection for efficient robust well placement optimization.
Water Cut Estimation in Electrical Submersible Pumps Using Artificial Neural Networks
An artificial lift is a method used to obtain a higher oil flow rate from the well, through some scheme that reduces the pressure at the bottomhole. Electrical submersible pumping is a common method in petroleum industry. The main component of this method is the electrical submersible pump (ESP), that can operate with complex flows involving mixtures of oil, water and gas. The presence of water in oil fields is a problem because it favors the formation of emulsions, which are the mixture of oil and water. Emulsions can be found in the form of oil-in-water and water-in-oil emulsions, depending on which phase is the continuous one and which is the dispersed one. Water-in-oil emulsions increase considerably the viscosity of the mixture and affect the pump’s efficiency, diminishing its pumping capacity. The increase or decrease of the water fraction in the process may cause the phenomenon called catastrophic phase inversion (CPI), in which the
dispersed phase becomes the continuous one and rapidly alters the physical properties of the flow, causing operational instability throughout the production system. In order to identify and predict this important phenomenon in complex multiphase flows, the usage of advanced identification tools, based on experimental data, has been used in recent years. In this work, artificial neural networks are used to estimate the water fraction in a flow that runs through an ESP. For that, data like inlet and outlet pressures, temperature, vibration and the correspondent water cut values, among others, were collected from an ESP operating with water and oil. Single-phase and two-phase tests were performed with the purpose of collecting data with different water cut values, ranging from 0% (single-phase oil) to 100% (two-phase water and oil). From the laboratory experiments, it was possible to build a data-driven computational tool capable of estimating the water fraction that runs through the pump, based on an optimized artificial neural network structure, which achieved an R-score of 0.9987.