The paper discusses the application of machine learning algorithms for one of the fields in Western Siberia. The results of ordinary attribute analysis and the results of the neural networks operation are presented. Their use is analyzed within the framework of a conceptual 3D geological model. The analysis and development of oil and gas fields with a complex clinoform structure is often complicated by the selection of a single seismic attribute for the entire area which allow to predict net thicknesses and reservoir properties in the interwell space (porosity, net thicknesses, etc.). To perform neural network calculation we used following data: seismic amplitude cube, structural surfaces that limit the search interval and participate in the generation of the low-frequency model. To train the neural network used logging curves and the values of the net thicknesses values at the well points. As a result of the calculations, the maps and 3D grid properties were generated to predict the selected parameter (net thickness) obtained in the variants P10, P50, P90, as well as the values of the standard deviation. The quality of the obtained results was assessed by the selecting optimal parameters of the algorithm, matching the results to conceptual representations and the actual wells operation data. As a part of the work we discuss a methodology for tuning machine learning algorithms and methods for assessing the quality of the described results. Comparison of ordinary and generated by neural networks attributes presented in paper. Results of the neural net calculations are presented by forecast maps of effective thickness (2D), grid properties (3D) of the West Siberian field. The advantage of using the described technique is confirmed by drilling new wells.
References
1. Priezzhev I.I., New age, Kolmogorov full functional neural network usage for nonlinear predictive seismic inversion, EAGE conference proceedings, Saint Petersburg, 2020, V. 2020, p.1 – 5, DOI: https://doi.org/10.3997/2214-4609.202053084
2. Kolmogorov A.N., Representation of continuous functions of several variables in the form of superpositions of continuous functions of one variable and addition (In Russ.), DAN, 1957, V. 114, pp. 953–956.
3. Kobrunov A. Priezzhev I., Hybrid combination genetic algorithm and controlled gradient method to train a neural network, Geophysics, 2016, V. 81(4), pp. 1–9, DOI:10.1190/geo2015-0297.1
4. Tikhonov A.N., Arsenin V.Ya., Metody resheniya nekorrektnykh zadach (Methods for solving ill-posed problems), Moscow: Nauka Publ., 1979, 223 p.
