- Home
- Conferences
- Conference Proceedings
- Conferences
Data Science in Oil & Gas
- Conference date: October 19-20, 2020
- Location: Online
- Published: 19 October 2020
1 - 20 of 24 results
-
-
Vertical variations of rock thermal conductivity from well-logging data using machine-learning methods
Authors A. Shakirov, Y. Meshalkin, D. Koroteev and Y. PopovSummaryThe problem of rocks thermal conductivity determination from well-logging data is highly important for many branches of petroleum geology and geophysics (basin modelling, thermal EOR modelling, thermal logging interpretation, etc.). Literature review revealed that there are only several studies performed to assess the applicability of neural network-based algorithms to predict rock thermal conductivity from well-logging data. In this research, we aim to define the most effective machine-learning algorithms for well-log based determination of rock thermal conductivity. Well-logging data acquired at a heavy oil reservoir together with results of thermal logging on cores extracted from two wells were the basis for our research. Eight different regression models were developed and tested to predict vertical variations of rock conductivity from well-logging data. Among considered machine-learning techniques, the Random Forest algorithm was found to be the most accurate at well-log based determination of rock thermal conductivity. From a comparison of the thermal conductivity depth profile predicted from well-logging data with the experimental data it can be concluded that thermal conductivity can be determined with a total relative uncertainty of 11.5%. The available data allows concluding that machine-learning algorithms are a promising framework for accurate well-log based predictions of rock thermal conductivity.
-
-
-
Method for increasing the resolution of seismic data based on convolutional neural networks
Authors T.V. Makeeva, N.D. Sergeeva, D.V. Egorov, T.I. Pashinskaya, O.V. Lukoyanycheva and I.A. TitaevaSummaryThe main problem of seismic research is the extremely low resolution of seismic data. Based on this, specialists need a tool that would be able to increase the resolution of seismic information. In this paper, we study application of convolutional neural networks with short-time Fourier transform to increase the resolution of seismic data. The approach was tested on synthetic seismic data and showed its effectiveness.
-
-
-
Machine learning in predicting discontinuous violations
Authors D. D. Sekerina, V. D. Gulin and G. S. GrigorievSummaryThis article is aimed at analysis the applicability of machine learning (ML) algorithms in predicting discontinuous violations based on potential field data. The text discusses common methods for identifying discontinuous violations and alternative methods based on machine logic using data from potential fields and their transformants. The advantages and disadvantages of algorithms in comparison with classical methods for determining discontinuous structures are also described. Using the example of one of the license areas (the East Siberian sea), we present a solution to the problem of fault mapping using machine prediction using ML and the “Reana” anomaly axis selection algorithm. In conclusion, we describe the features of working with the methods of automatic prediction of discontinuous violations, the relevance of further development of the application of this direction. It is shown that the algorithm is efficient and suitable for practical use.
-
-
-
Seismostratigraphic segmentation using machine learning. Eastern Siberia case study.
Authors A.A. Koziaev, A.S. Tsoy and G.V. IvanovSummaryThe article presents the results of the use of machine learning methods in the interpretation of seismic data obtained on the land of the Russian Federation, obtained in the complex seismic-geological conditions of Eastern Siberia. Seismostratigraphic segmentation into 7 zones using convolutional neural networks has been performed. For training the algorithm, 10% of the total amount of data is used. Comparison of the ML-based algorithm with standard approaches to auto-tracking is carried out. The results obtained allow us to speak about the high efficiency of machine learning methods for interpreting seismic data obtained in complex seismic-geological conditions of Eastern Siberia. In addition, it should be noted that algorithms based on the effective use of ML with “non-phase” correlation of their information for horizons confined to layers with a complex geological structure (reefs, channels, etc.), while standard algorithms for auto-tracking horizons do not cope with this task.
-
-
-
Inversion, neural networks, test construction and logging interpretation problems
Authors B.N. Enikeev and L.R. FattakhovSummaryThe report discusses the expediency of using full-fledged synthetic tests to solve problems associated with intellectual data analysis in the oil and gas industry. Several main areas of application of such tests are considered: for corporate training of interpreters to work with different algorithms for evaluation of complex reservoirs; for preliminary training of such algorithms under conditions of limited field test information; for determination of the area of applicability of these algorithms, including forecasting tasks. The concept of preparation of realistic tests, both for main section types and for log data complexes, is implemented within the framework of system approach. This approach is based on the ideology of “statistical interpretation” according to the known concept of forward and reverse tasks. The software and algorithmic implementation of test construction procedures is based on the knowledge of petrophysical relationships, as well as on practical experience of its application and correction in conditions of different section types. The practical use of the test system is illustrated by the example of comparison of the relative efficiency of the algorithms depending on the volume of training materials, their noisiness and orientation to the interpolation or forecast task. The work performed shows the necessity to solve various questions (including petrophysical, organizational, methodological, and protocol) with the involvement of a wide range of professionals of different profiles.
-
-
-
Development of an algorithm for analyzing and integrating the results of different-scale studies of reservoir filtration properties
By N.D. KozyrevSummaryThe hydrodynamic model allows you to predict many technological indicators for the short and long term, such as oil and gas production, watering rates, reservoir and bottom-hole pressures and many others. One of the main advantages of using geological-hydrodynamic models in predicting of field development is the consideration of geological heterogeneity, which allows to accurately predict the filtration and physicochemical processes occurring in reservoirs. The reliability and accuracy of the forecast done with the hydrodynamic model directly depends on the quality and quantity of the initial information taken into account in creating the model. The assessment and accounting of the results of various methods of reservoir analysis is often carried out separately from each other and at different times of model creation. Using this approach, in particular for deposits with a complex geological structure, the degree of uncertainty of the reservoir properties remains high, even with a sufficient information, which negatively affects the technological and economic projections of field development. Therefore, to ensure the high quality of the predictive ability of the hydrodynamic model, a comprehensive analysis of all initial information is necessary
-
-
-
Using machine learning algorithms for the interpretation the results of seismic data processing using the RTH method on the example of a field in Eastern Siberia
Authors R.V. Simonov and E.V. AnokhinaSummaryThe development of digital algorithms that allow you to automate certain interpretation tasks makes it much easier to solve geological problems. The main difficulty of automatic algorithms in complex geological conditions is related to the multivariance of interpretation and, as a result, possible errors that can introduce ambiguity in further interpretation. The use of machine learning methods and Reverse Time Holography can help in solving this problem, as well as reduce the amount of work performed manually by the interpreter and improve the accuracy of the results.
-
-
-
Well-Logging based lithology prediction using Machine Learning
Authors Y. Meshalkin, A. Shakirov, D. Orlov and D. KoroteevSummaryIn petroleum systems, geological research studies include the classification of formation lithology by using weblogging methods. With the development of computer technology and decision-making systems based on machine learning methods, interest to automate lithology classification tasks from well logging is growing. This is especially important when manually processing large amounts of information. In this paper, we aim to define the most effective machine learning techniques for well log-based determination of lithology on the example of oil field in western Siberia, Russia. There are 86 wells each of which contains 6 physical parameters: electrical resistivity, gamma-ray, density, well diameter, photoelectric factor and neutron porosity. We applied a Linear model, k-nearest neighbors, Artificial Neural Network, Gradient Boosting, Random Forest, AdaBoost, and recently developed gradient boosted decision tree (GBDT) systems, namely, XGBoost, LightGBM, and CatBoost on real data. The performance of algorithms was compared and evaluated using metrics such as accuracy and macro F1-score (F1) on the test set after hyperparameter tuning. As a result, among the applied algorithms, we found that CatBoost possessed the highest metrics. In our work, we have demonstrated the high generalizing ability of Machine learning methods within the chosen research object for well-logging based lithology classification problem.
-
-
-
Using neural networks for geophysical data compression
By V.I. BulaevSummaryIn this paper, an approach for lossy geophysical data compression is considered. Principles of using neural network and wavelet-transform for compression system is discussed, which confirms high efficiency of proposed approach.
-
-
-
Automation of LWD interpretation
Authors A.S. Romantsov, A.M. Galechyan and V.V. KapsenkovSummaryDrilling process geological monitoring becomes more critical due to horizontal drilling percentage increase over the last two decades. That results in a growth of LWD operative interpretation demand. This paper represents the automation of LWD interpretation process. Decreasing the process life of operative interpretation we manage to increase the geosteering efficiency. This work examines formation evaluation and calculation of the mineral volumetric fractions. Collector, lithology and formation fluid type are determined by the cut-off values method. Data transmission is performed via the classified remote server or by manual import and export via the email. The represented approach has shown its efficiency on practice. It makes log quality control more effective and optimizes the workflow by cutting down unproductive routine operations.
-
-
-
Data analysis and numerical modeling for forecasting, prospecting and exploration of oil and gas
Authors V.Yu. Kerimov and R.N. MustaevSummaryThe article describes the importance of data analysis and numerical modeling in forecasting, prospecting and exploration of oil and gas fields. Numerical modeling of hydrocarbon systems is carried out at different stages of exploration work by applying a series of projects aimed at implementing the tasks of individual stages and stages, based on data analysis and the development of new, and adaptation of existing technologies.
-
-
-
Well log clustering as a tool for rock typing of carbonate reservoir
Authors I.I. Churochkin, P.A. Kharitontseva and A.A. RoslinSummaryThe carbonate reservoir is one of the most “inconvenient” and challenging reservoir types that a geologist encounters in his daily work. The complex pore space of the carbonate reservoir is associated with secondary porosity, the appearance of which is facilitated by processes such as leaching, dolomitization, recrystallization and fracturing. Due to these processes, the primary matrix is transformed, and the reservoir properties are no longer controlled by the lithofacies. In other words, rocks deposited in similar geological conditions can have completely different reservoir properties. This paper presents the results of testing the hypothesis for identifying rock types by clustering well logging and geological substantiation of the identified groups. At the first stage, clustering algorithms helped to divide the values of the curves into groups with specific responses. Further, the data on the lithological description of the core, capillary pressure curves, as well as curves of relative permeabilities made it possible to substantiate the correctness of the identified clusters.
-
-
-
Application of genetic algorithms for multimineral modeling based on the principle of petrophysical inversion
Authors V. Rudenko, I. Babakov and I. PriezzhevSummaryOne of the necessary steps of well-log interpretation is a multimineral modeling. Unconventional and complex reservoirs require a special approach to their search and exploration because standard applied methods are not always characterized by good results. As a solution to the problem of unconventional reservoirs studying we propose application of non-linear stochastic algorithms for multimineral modeling based on the principle of petrophysical inversion. The developed algorithm allows calculating changes in the physical properties of minerals at each point in depth. Calculated logs of tool response can be an additional tool of analysis for assessing the prospects of unconventional and complex reservoirs.
-
-
-
Automated construction of a spatial hydrogeochemical model of Jurassic-Cretaceous sediments of the West Siberian megabasin
Authors A.G. Plavnik, M.V. Itskovich and V.P. AstafievSummaryThe paper presents a method and software for three-dimensional modeling of the groundwater chemical composition, constructing cross-sections and maps of groundwater chemical parameters detailed up to individual aquifers with functionality for assessing the results’ reliability. The developed software module provides efficient processing of large volumes of hydrogeochemical information and automated calculations for a complex of groundwater components.
-
-
-
Interpretation of seismic inversion results using the “Random Forest”
By A.V. ButorinSummaryThe study is aimed to estimate the possibility of using machine-learning method “Random Forest”, to obtain a probabilistic estimate of the distribution of an oil-saturated reservoir. The object of research is the Achimov complex, composed of relatively thin interlayered terrigenous rocks. The “random forest” method realized with the scikit-learn Python library of the. Application of the algorithm converts the input cubes of elastic parameters to probability cubes of lithotypes, which used for geological interpretation. As a result, the trends of reservoir properties estimated, as well as the probability cube of an oil-saturated reservoir. These data can be effectively used in planning the well-paths.
-
-
-
Metric for evaluating difference between seismic gathers
Authors A.I. Arefina, R.L. Khudorozhkov, A.S. Kuvaev and D.G. SeminSummaryWhen building machine learning (ML) models for specific tasks having a good metric is of crucial importance. Constructing a good metric that is relevant to the problem domain is a difficult task. We consider a problem of noise attenuation on seismic data, specifically ground roll attenuation problem. The methods of quality control that are used by experts who perform the denoising task manually can not be directly transferred to machine learning setting. A generalization of such a method is proposed to obtain a formal computable metric that allows to compare denoised gathers with respect to their “proximity” to a ground truth cleaned data, and consequently to compare ML models for ground roll attenuation.
-
-
-
Machine learning models for Prognostic Health Management in Oil and Gas industry
Authors D. Bannikov, D. Pantsurkin, I. Velikanov, K. Lyapuunov and S. ParkhonykSummaryIn this work, we present an approach to enable prognostic health management for fracturing equipment by implementing failure identification and prediction models. It combines both domain knowledge (failure modes of the equipment, sensors data behavior over time) and machine learning (time series analysis, data preprocessing, and feature engineering)
-
-
-
Applying Machine Learning Methods to Identify Injection Intervals Using DTS Data
Authors A.A. Sadretdinov, R.A. Valiullin and R.K. YarullinSummaryThe problem of determining the intervals of fluid loss in an injection horizontal well is solved. The chosen setting is one of the simplest from the point of view of the ongoing processes. The successful solution of the problem using machine learning methods should show the perspective of the approach for solving more complex problems. The chosen approach to solving the problem involves training models on a synthetic sample obtained using a thermohydrodynamic simulator. The results of the algorithm operation are demonstrated, which can be assessed as successful.
-
-
-
Method of focusing neural network approximators for solving a nonlinear inverse problem of geophysics
Authors I.E. Obornev, M.I. Shimelevich, E.A. Obornev and E.A. RodionovSummaryNeural networks (NN) are widely used to solve various problems of interpretation and processing of geophysical data. In this paper, we consider the application of the approximation neural network (ANN) method for solving inverse problems (including multicriteria), which are reduced to a nonlinear operator equation of the first kind of the form (respectively, to a system of operator equations). ANN method consists of constructing an approximate inverse operator of the problem using neural network approximation constructions (MLP networks) based on a set of support solutions of direct and inverse problems built in advance. Surface inhomogeneities of the medium create a significant noise component in determining the parameters of the underlying regions, which leads to a significant increase in the error in solving the inverse problem. The proposed method of focusing NN approximators consists in the fact that several different approximators are constructed based on several training sets, which differ in the detail of the parameterization of the studied environment at different depths. Focusing NN approximators can achieve a reduction of its training error for environment parameters, which focuses on approximator. This allows one to significantly reduce the intrinsic errors of the NN method when solving nonlinear inverse 3D problems of geophysics.
-
-
-
Integration of all geological and geophysical information and construction of a geological model using the geostatistical inversion algorithm
More LessSummaryThe main task of any geological and geophysical research is to build a unified geological model by combining consistent information from all available sources. Proper integration of well, geological and seismic information is a key task, the solution of which allows you to get a single model that does not contradict any of the sources of initial information. Traditionally, incorporation of data is carried out after receiving the results of individual studies or not at all, preferring one of the methods of data interpretation. This paper will consider a geostatistical approach that allows you to combine information from various sources directly when building a geological model.
-