Intellectual analysis as a method of knowledge discovery in field development

UDK: 004.0322.26:622.276.1/.4

DOI: 10.24887/0028-2448-2025-5-132-136

Key words: intellectual analysis, neural networks, multilayer perceptron, multistage hydraulic fracturing, regression, forecasting, field development

Authors: V.A. Markin (Surgutneftegas PJSC, RF, Surgut); L.V. Markina (Oil and Gas Production Department Fedorovskneft, Surgutneftegas PJSC, RF, Surgut); V.R. Bayramov (Surgutneftegas PJSC, RF, Surgut); M.Yu. Lobanok (Surgutneftegas PJSC, RF, Surgut); E.D. Shvechikov (SurgutNIPIneft, Surgutneftegas PJSC, RF, Tyumen); D.R. Ayupov (SurgutNIPIneft, Surgutneftegas PJSC, RF, Tyumen); E.G. Bushmeleva (SurgutNIPIneft, Surgutneftegas PJSC, RF, Tyumen)

The article is dedicated to the issue of creating predictive models based on trained user-defined and automated neural networks for forecasting certain production characteristics of horizontal wells with multistage hydraulic fracturing. According to the authors, the predicted characteristics are fundamental in assessing the potential of oil wells or the effectiveness of well intervention techniques. A numerical prediction (regression) task was defined and solved, a comprehensive approach to training both user-defined and automated neural networks is presented, the architecture and free parameters of neural networks are experimentally determined, and an optimal set of input data for modeling is identified using the «backwards elimination» method, which is often applied in statistics but rarely used with neural networks. It is noted that the process of training neural network is largely hidden and remains unexplained (which is why neural networks have a reputation as a «black box»). In turn, the conducted research demonstrates criteria for selecting the most accurate predictive models, assessing the importance of variables for analysis, and tools for evaluating model outputs, which significantly unveils the «black box» of the neural network process. Based on the geological nature of the studied object, the possibility of replicating trained predictive models is demonstrated, without being limited to a single subsurface area. Thus, the authors propose implementing neural network predictive models capable of correctly forecasting well production characteristics under conditions of significant data variability and heterogeneity typical of many operations in field development.

References

1. Nisbet R., Elder J., Miner G., Handbook of statistical analysis and data mining applications, Academic Press, 2009, 822 p., DOI: https://doi.org/10.1016/B978-0-12-374765-5.X0001-0

2. Neyronnye seti. STATISTICA Neural Networks: Metodologiya i tekhnologii sovremennogo analiza dannykh (Neural networks. STATISTICA Neural Networks: Methodology and technologies of modern data analysis): edited by Borovikov V.P., Moscow: Goryachaya liniya – Telekom Publ., 2008, 392 p.

3. Markin V.A., Markina L.V., Bayramov V.R., Lobanok M.Yu., Data Mining methods as a decision support system under conditions of data limitation (In Russ.), Neftyanoe khozyaystvo = Oil Industry, 2024, no. 5, pp. 138–142, DOI: https://doi.org/10.24887/0028-2448-2024-5-138-142