METHOD FOR REGIONAL FORECAST OF OIL AND GAS POTENTIAL TERRITORIES BY MACHINE LEARNING ALGORITHMS ON THE EXAMPLE OF THE TYUMEN FORMATION OF WESTERN SIBERIA

• Published in: Izvestiya Tomskogo Politekhnicheskogo Universiteta Inziniring Georesursov
• Main Author: Dmitry A. Ivlev
• Format: Article in Journal/Newspaper
• Language: Russian
• Published: Tomsk Polytechnic University 2021
• Subjects: tyumen formation; middle jurassic; machine learning; artificial intelligence; feature generation; feature selection; gradient boosting; generative adversarial neural; networks; regional oil and gas forecast; probability of geological success; hydrocarbon resource; geoinformatic; basin modeling; sedimentation modeling; modeling hc migration; Engineering geology. Rock mechanics. Soil mechanics. Underground construction; TA703-712; khanty; nenets; Yamalo Nenets; Siberia
• Online Access: https://doi.org/10.18799/24131830/2021/1/2998; https://doaj.org/article/b4e8470f09014b70a00a2326d2368d98

The relevance of research is caused by the reduction in the fund of structural traps and the need to expand the resource base of hydrocarbons by increasing the efficiency of prospecting and exploration of fields in complex oil and gas deposits. The main aim of the research is to show the forecasting methodology and the set of applied technological solutions and algorithms using the example of forecasting the oil and gas content of the study area. Object: Middle Jurassic deposits (Tyumen Formation) of Western Siberia within the region (700×900 km), which includes parts of the Yamalo-Nenets and Khanty-Mansiysk administrative districts and the Tomsk region. Methods. Using the machine-learning algorithms and integrating a technological set of methods: geoinformatics, basin modeling, and expert assessments the following stages of the forecast method implementation are shown: 1) generation of the feature space of the studied area based on increasing the spatial resolution of structural constructions using algorithms of generative-adversarial architecture of neural networks, where the results of 3D seismic survey are used as reference areas; 2) selection of features by statistical method and machine learning methods; 3) creation of a subset of forecast models based on gradient boosting over decision trees; 4) combining them into a metamodel by stacking generalization by logistic regression. Results. An approach to regional forecasting has been formalized and tested. A forecast of the probability of oil and gas content of the Tyumen suite in the study area was made. On its basis and information on discovered fields, the hydrocarbon resource base was estimated by the Monte Carlo method. The results are presented in the form of a summary table of geological and recoverable resources for probabilities P10, P50, P90 in comparison with the categories of reserves ABC1 and ABC1+C2 of the fields listed on the state balance sheet in the study area. As an example, the graphic materials of the results are given: the work of the ...