Многомасштабное моделирование нефтегазовых процессов на основе активного обучения суррогатных моделей тема диссертации и автореферата по ВАК РФ 00.00.00, кандидат наук Печко Константин Анатольевич

  • Печко Константин Анатольевич
  • кандидат науккандидат наук
  • 2025, «Национальный исследовательский университет ИТМО»
  • Специальность ВАК РФ00.00.00
  • Количество страниц 248
Печко Константин Анатольевич. Многомасштабное моделирование нефтегазовых процессов на основе активного обучения суррогатных моделей: дис. кандидат наук: 00.00.00 - Другие cпециальности. «Национальный исследовательский университет ИТМО». 2025. 248 с.

Оглавление диссертации кандидат наук Печко Константин Анатольевич

Реферат

Synopsis

Введение

ГЛАВА 1. Обзор методов моделирования нефтегазовых

процессов

1.1 Подходы к математическому моделированию нефтегазовых процессов

1.1.1 Физико-математические модели

1.1.2 Эмпирические модели

1.1.3 Статистические и вероятностные модели

1.1.4 Модели машинного обучения

1.2 Интегрированное моделирование актива как современный

подход моделирования нефтегазовых процессов

1.2.1 Концепция замкнутого цикла управления

1.2.2 Преимущества интегрированного моделирования

1.2.3 Ограничения и вызовы

1.3 Оптимизация с ресурсными ограничениями в интегрированном моделировании нефтегазовых активов

1.3.1 Постановка задачи

1.3.2 Подходы к решению

1.3.3 Актуальные решения

1.4 Суррогатное моделирование нефтегазовых процессов

1.4.1 Основы суррогатного моделирования

1.4.2 Преимущества и ограничения суррогатных моделей

1.4.3 Применение суррогатных моделей в задачах нефтегазовой индустрии

1.5 Основы активного обучения

1.5.1 Постановка задачи активного обучения

1.5.2 Стратегии активного обучения

1.6 Выводы по главе

ГЛАВА 2. Методология многомасштабного моделирования нефтегазовых процессов на основе активного

обучения суррогатных моделей

2.1 Метод и алгоритмы многомасштабного моделирования нефтегазовых проессов

2.1.1 Формализация многомасштабного моделирования нефтегазовых систем

2.1.2 Компонентная декомпозиция ИМА

2.1.3 Графовое представление ИМА

2.1.4 Сопряжение моделей различной детальности

2.1.5 Стратегия перехода между уровнями моделирования

2.1.6 Архитектура многомасштабной симуляции ИМА

2.2 Метод и алгоритмы построения суррогатных моделей-компонент ИМА

2.2.1 Этапы суррогатного моделирования

2.2.2 Суррогатная модель пласта

2.2.3 Суррогатная модель скважины

2.2.4 Суррогатная модель наземной инфраструктуры

2.2.5 Прокси-ИМА и ее преимущества

2.3 Метод и алгоритмы активного обучения суррогатных моделей

для многомасштабного моделирования нефтегазовых процессов

2.3.1 Стратегии активного обучения

2.3.2 Псевдокод алгоритма активного обучения

2.3.3 Схема цикла активного обучения

2.4 Выводы по главе

ГЛАВА 3. Программная реализация методологии

многомасштабного моделирования на основе активного обучения суррогатных моделей

3.1 Архитектура программного комплекса

3.1.1 Общая структура и принципы организации программного комплекса

3.1.2 Взаимодействие модулей и принципы интеграции

3.1.3 Технологический стек и расширяемость

3.2 Суррогатное моделирование компонентов ИМА

3.2.1 Реализация суррогатной модели пласта

3.2.2 Реализация суррогатной модели скважины

3.2.3 Реализация суррогатной модели инфраструктуры

3.2.4 Интеграция компонентов и формирование прокси-ИМА

3.3 Реализация процедур активного обучения

3.3.1 Архитектурные основы модуля активного обучения

3.3.2 Метрики качества и критерии отбора

3.3.3 Стратегии выбора новых точек

3.3.4 Гибкость и масштабируемость подхода

3.3.5 Интеграция с системой логирования и воспроизводимость

3.4 Сопряжение моделей различной детальности

3.5 Выводы по главе

ГЛАВА 4. Практическое применение и внедрение разработок

4.1 Эксперементы на синтетических данных

4.1.1 Эксперементальная апробация методологии многомасштабного моделирования

4.1.2 Экспериментальная апробация методологии активного обучения суррогатных моделей

4.2 Валидация на реальных задачах нефтегазовой отрасли

4.2.1 Решение задачи реновации трубопроводов системы сбора

и транспорта

4.2.2 Оптимизация базового фонда месторождения

4.3 Внедрение полученных результатов

Заключение

Список сокращений и условных обозначений

Словарь терминов

Список литературы

Список литературы

Приложение

Тексты публикаций

Реферат

Общая характеристика диссертации

Рекомендованный список диссертаций по специальности «Другие cпециальности», 00.00.00 шифр ВАК

Введение диссертации (часть автореферата) на тему «Многомасштабное моделирование нефтегазовых процессов на основе активного обучения суррогатных моделей»

Актуальность исследования

Современная нефтегазовая отрасль требует высокоточного и оперативного прогнозирования производственных процессов в условиях сложных, неоднородных данных и жестких временных ограничений. Классические методы математического моделирования, основанные на физических законах, обеспечивают высокую детализацию, но сталкиваются с чрезмерными вычислительными затратами, особенно при учете множества взаимосвязанных факторов. Это делает актуальным поиск решений, сочетающих точность физических моделей с вычислительной эффективностью машинного обучения (МО).

Многомасштабное моделирование, интегрирующее детальные физические модели и суррогатные аппроксимации на основе МО, позволяет значительно ускорить расчеты без потери точности. Однако существующие подходы зачастую не обеспечивают эффективной адаптации суррогатных моделей к изменяющимся условиям, требуя ручной настройки и значительных вычислительных ресурсов для переобучения.

Активное обучение суррогатных моделей открывает новые возможности в этом направлении, позволяя автоматически выбирать наиболее информативные данные для обучения, минимизируя объем вычислений и повышая точность прогнозирования. Такой подход особенно важен для задач оперативного управления добычей, где требуется быстрая адаптация моделей к новым данным.

Разработка методов многомасштабного моделирования на основе активного обучения суррогатных моделей позволит создать более эффективные цифровые двойники месторождений, обеспечивающие высокую скорость и точность расчетов.

Степень разработанности темы

Вопросы моделирования разработки нефтегазовых месторождений с использованием методов математического и компьютерного моделирования

интенсивно изучаются как в отечественной, так и в зарубежной научной литературе. Традиционно основное внимание уделялось построению высокоточных цифровых двойников, базирующихся на численном решении систем дифференциальных уравнений. Такие модели позволяют достичь высокой точности при прогнозе, но сопровождаются существенными вычислительными затратами, особенно в условиях многовариантного анализа или необходимости оперативного принятия решений.

В последние два десятилетия активно развивается направление суррогатного моделирования (surrogate modeling), направленное на сокращение затрат путём замены ресурсоёмких численных симуляций более лёгкими в вычислительном плане аппроксимациями. В международной практике фундаментальными работами в данной области являются исследования A. I. Forrester и соавт. [1], в которых предложены подходы к построению многомасштабных (multi-fidelity) суррогатных моделей. Подходы Forrester стали основой для большого числа прикладных работ в области аэрокосмической, автомобильной и энергетической промышленности, где требуется оптимизация по множеству сценариев.

Другим важным направлением является многомасштабное моделирование, предполагающее совместное использование моделей различной степени детальности (например, грубосеточных и высокоточных моделей) в рамках единого вычислительного процесса. Концепция сопряжения моделей разных уровней используется в задачах оптимизации добычи, планирования разработки, инверсии параметров пластов. В области количественной оценки неопределённостей (UQ), значительный вклад внесён группой Y. Marzouk (MIT), развивающей методы построения и калибровки суррогатов с учётом информации разных уровней точности. [2].

В контексте активного обучения (active learning), актуального при ограниченном бюджете на получение обучающих данных, развиваются подходы, позволяющие целенаправленно выбирать наиболее информативные точки симуляции. Эти подходы находят применение в инверсии, планировании экспериментов и обучении метамоделей. Методологическая база таких методов включает байесовский оптимизационный анализ, оценку неопределённости предсказаний и

теорию информационного планирования. В российской научной среде вклад в развитие соответствующих методов внесён, в частности, К. В. Воронцовым, развивающим направление активного обучения и выбора оптимальных подмножеств данных для задач регрессии и классификации. Хотя его исследования преимущественно ориентированы на обобщённые задачи машинного обучения, предложенные им методологические принципы находят применение в инженерных задачах с ограниченными ресурсами.

В прикладной нефтегазовой области отечественными исследователями предпринимались попытки учитывать иерархии моделей и масштабные эффекты. В частности, М. М. Хасанов в своих работах подчёркивал важность учёта масштабной зависимости параметров и последовательной калибровки моделей различной физической детализации. Эти идеи находят продолжение в построении цифровых двойников, где необходимо интегрировать данные и модели на разных уровнях.

Следует также отметить, что методологическая база построения суррогатных моделей и активного обучения активно развивается в ряде отечественных научных школ. В частности, Сколковский институт науки и технологий (Skoltech) и Университет ИТМО ведут исследования в области машинного обучения , что составляет основу для построения суррогатных моделей инженерных систем.

Несмотря на наличие значительного количества научных и прикладных работ, на текущий момент отсутствуют комплексные решения, обеспечивающие автоматизированное активное обучение суррогатных моделей в многомасштабной вычислительной среде применительно к задачам оперативного моделирования и оптимизации разработки нефтегазовых месторождений. Существующие решения либо ограничиваются использованием суррогатов в фиксированных условиях, либо не обеспечивают эффективную интеграцию с физическими моделями и не решают задачу динамической адаптации при изменении производственной информации.

Таким образом, представленная работа направлена на устранение выявленного научно-практического пробела за счёт разработки методов гибридного

многомасштабного моделирования с использованием активного обучения, позволяющих обеспечить как высокую точность, так и вычислительную эффективность в задачах моделирования процессов в нефтегазовой отрасли.

Объект исследования — процесс моделирования разработки нефтегазовых месторождений с использованием методов математического и компьютерного моделирования.

Предмет исследования — методы, алгоритмы и программное обеспечение гибридного моделирования нефтегазовых процессов на основе активного обучения суррогатных моделей.

Цель исследования — разработка и внедрение методов многомасштабного моделирования нефтегазовых процессов с использованием активного обучения суррогатных моделей, обеспечивающих возможность оперативного анализа большего числа сценариев разработки месторождений и повышение обоснованности принимаемых решений в условиях ограниченности времени и вычислительных ресурсов, с сохранением требуемой точности.

Для достижения поставленной цели необходимо было решить следующие задачи исследования:

1. Разработать метод и алгоритмы многомасштабного моделирования нефтегазовых процессов, обеспечивающие сопряжение моделей различной детальности в рамках интегрированной модели актива.

2. Разработать метод и алгоритмы построения суррогатных моделей нефтегазовых процессов на основе машинного обучения, ориентированные на эффективную аппроксимацию результатов численного моделирования.

3. Разработать метод и алгоритмы активного обучения суррогатных моделей, позволяющие минимизировать количество дорогостоящих симуляций за счет выбора наиболее информативных точек обучения.

4. Программно реализовать предложенные методы и алгоритмы, а также провести их верификацию на реальных и приближенных к реальным задачах моделирования нефтегазовых процессов.

Научная новизна исследования заключается в разработанных методах многомасштабного моделирования нефтегазовых процессов, основанных на

интеграции детальных физических моделей с суррогатными аппроксимациями, построенными с использованием активного обучения. Впервые предложен подход, обеспечивающий автоматизированный выбор информативных точек для обучения суррогатных моделей, в рамках интегрированного моделирования активов. Это позволяет существенно снизить вычислительные затраты при сохранении высокой точности моделирования. Реализован программный комплекс, позволяющий применять разработанные методы и алгоритмы к актуальным задачам нефтегазовой отрасли. Продемонстрирована его эффективность на синтетических и реальных тестах.

Теоретическая значимость работы Теоретическая значимость исследования заключается в развитии научных подходов к моделированию сложных инженерных задач в условиях ограниченных ресурсов. Полученные результаты расширяют фундаментальные основы построения эффективных вычислительных схем для моделирования и оптимизации процессов разработки нефтегазовых месторождений, что способствует дальнейшему развитию теории многомасштабных и гибридных вычислительных методов в отрасли.

Практическая значимость работы заключается в разработке и внедрении эффективных методов и программных средств, позволяющих существенно снизить вычислительные затраты и повысить оперативность анализа при моделировании и оптимизации разработки нефтегазовых месторождений. Разработанные алгоритмы гибридного многомасштабного моделирования и активного обучения суррогатных моделей интегрированы в корпоративные цифровые платформы российских компаний. Разработанный программный комплекс обеспечивает повышение точности и скорости инженерных расчетов, способствует более обоснованному принятию решений и повышает эффективность эксплуатации нефтегазовых активов в реальных производственных задачах и уже используется инженерами.

Методы исследования: методы математического и компьютерного моделирования, машинного обучения, оптимизации, математической статистики и программной инженерии.

Положения выносимые на защиту.

1. Методология и алгоритмы суррогатного моделирования нефтегазовых процессов на основе методов машинного обучения, обеспечивающие снижение вычислительных затрат при сохранении требуемой точности моделирования.

2. Методология и алгоритмы гибридного многомасштабного моделирования, предусматривающие интеграцию детальных физических моделей и их суррогатных аппроксимаций с использованием стратегий активного обучения для оптимального распределения вычислительных ресурсов.

3. Программная реализация разработанных методов и алгоритмов для многомасштабного моделирования нефтегазовых процессов, обеспечивающая автоматизированное построение и сопряжение моделей различной детальности, а также верификацию разработанных методов на синтетических и реальных промышленных задачах.

Достоверность научных положений, выводов и практических рекомендаций, полученных в рамках данной диссертационной работы, подтверждается:

— Эмпирическим подходом при разработке методов и алгоритмов;

— Верификацией на синтетических и реальных данных, сопоставлением результатов суррогатных моделей с численными расчетами и промысловыми данными;

— Апробацией в опытно-промышленных испытаниях, успешным внедрением методологии и программного комплекса в производственные процессы группы команий «Газпромнефть», что подтверждает практическую применимость и валидность предложенных решений;

— Тестированием с альтернативными методами — демонстрацией преимуществ разработанного подхода по точности и вычислительной эффективности;

— Публикациями в рецензируемых журналах, докладами на международных конференциях, полученными патентами и свидетельствами — независимой экспертной оценкой результатов в научном сообществе.

Апробация работы: Результаты исследования были представлены на следующих конференциях:

— Конференция «Интеллектуальный анализ данных в нефтегазовой отрасли» (2022, Новосибирск, Россия);

— Конференция «Геомодель 2022» (2022, Геленджик, Россия);

— Конференция «Conference on Digital Innovation for a Sustainable Future» (2022, Бангкок, Таиланд);

— Научно-техническая конференция «НТК Газпромнефть» (2022, Санкт-Петербург, Россия);

— Конференция «EAGE Digital» (2023, Лондон, Великобритания);

— Конференция «IFEDC» (2023, Ухань, Китай);

— Научно-техническая конференция «НТК Газпромнефть» (2023, Санкт-Петербург, Россия);

— Форум «Молодые Профессионалы» (2023, Санкт-Петербург, Россия);

— Конференция «Fourth EAGE Digitalization Conference & Exhibition» (2024, Париж, Франция);

— Конференция «85th EAGE Annual Conference & Exhibition» (2024, Осло, Норвегия);

— Научно-техническая конференция «НТК Газпромнефть» (2024, Санкт-Петербург, Россия);

— Конференция «МГРиМО» (2024, Санкт-Петербург, Россия);

— Конференция «Интеллектуальный анализ данных в нефтегазовой отрасли» (2024, Калининград, Россия);

— Конференция «Digital Twins and Predictive Analytics» (2024, Куала-Лумпур, Малайзия);

— Конференция «EAGE Digital» (2025, Эдинбург, Шотландия).

Проведены опытно-промышленные работы на моделях и активах компании

«Газпромнефть».

Личный вклад: Автор принимал активное участие во всех этапах исследования, включая разработку и внедрение методов и алгоритмов многомасштабного моделирования нефтегазовых процессов, построение и тестирование

суррогатных моделей, анализ и предобработку данных, а также формулирование методологии использования машинного обучения в ИМА. Автор активно участвовал в оформлении научных публикаций и патентных заявок. Лично разработал методики построения суррогатных моделей для задач многомасштабного моделирования. Провел анализ существующих методов математического моделирования и машинного обучения, выполнил адаптацию алгоритмов для задач нефтегазовой отрасли. Разработал программный комплекс, реализующий предложенные методы, включая написание кода, тестирование и документирование. Лично выполнил верификацию и валидацию предложенных методов и алгоритмов на реальных и синтетических данных. Провел обработку и интерпретацию результатов численных экспериментов. Участвовал в апробации разработок на реальных задачах в сотрудничестве с инженерами нефтегазовой отрасли.

Публикации: Основные результаты по теме диссертации изложены в 16 печатных изданиях, 6 из которых изданы в журналах, рекомендованных ВАК, 4 в журналах Scopus или Web of Science, 6 — в тезисах докладов. Зарегестрирован 1 патент на изобретение, 2 программы для ЭВМ.

Публикации в научных изданиях, входящих в международные реферативные базы данных и системы цитирования:

1. Voloskov D. et al. Production rate forecasting using pressure and saturation estimates near the wellbore // Geoenergy Science and Engineering. - 2025. -P. 213946.

2. Pechko K. et al. Leveraging Model Predictive Control for Intelligent Field Development Scenario Generation in Oil and Gas // Fifth EAGE Digitalization Conference & Exhibition. - European Association of Geoscientists & Engineers, 2025. - Vol. 2025. - No. 1. - P. 1-3.

3. Pechko K.A., Kim V.V., Matroshilov N., Simonov M., Afanasev A. Methodology for selecting analogs of reservoir fluid PVT models and rapid estimation of PVT parameters for new assets (Russian) // Neftyanoe khozyaystvo-Oil Industry. - 2024. - No. 12. - P. 36-39. (Q3)

4. Kim V. et al. A New Approach to Rapid Generation of PVT Black-Oil Model Using Machine Learning Techniques for Greenfields // European Association of Geoscientists & Engineers. - 2024. - Vol. 2024. - No. 1. - P. 1-5.

5. Pechko K. et al. Data Pre Processing Techniques in Integrated Asset Modeling // EAGE Digitalization. - 2024. - Vol. 2024. - No. 1. - P. 1-4.

6. Pechko K. et al. Optimizing Oil and Gas Assets Using Machine Learning Algorithms at Different Levels of Detail // Energy Excellence: Digital Twins and Predictive Analytics. - 2024. - Vol. 2024. - No. 1. - P. 1-3.

7. Pechko K. et al. Application of Machine Learning in Integrated Modeling of the Oil and Gas Fields // EAGE Digitalization. - 2023. - Vol. 2023. - No. 1.

- P. 1-4.

8. Kim V.V. et al. Methodology for selecting analogs of reservoir fluid PVT models and rapid estimation of PVT parameters for new assets // Neftyanoe khozyaystvo-Oil Industry. - 2023. - Vol. 2023. - No. 12. - P. 36-39.

9. Pechko K. et al. Acceleration of Calculations of Integrated Fields Using Machine Learning Methods // EAGE Conference on Digital Innovation for a Sustainable Future. - European Association of Geoscientists & Engineers, 2022. - Vol. 2022. - No. 1. - P. 1-4.

10. Pechko K.A., Senkin I.S., Belonogov E.V. Modeling of Two-Phase Fluid Flow in a Well Using Machine Learning Algorithms // Data Science in Oil and Gas 2021. - European Association of Geoscientists & Engineers, 2021. -Vol. 2021. - No. 1. - P. 1-5.

Публикации в научных изданиях, входящих в перечень российских рецензируемых журналов:

1. Печко К. А. и др. Факторный анализ в экономическом модуле интегрированной модели актива // РЯОНЕФТЬ. Профессионально о нефти.

- 2024. - Т. 9. - №. 2. - С. 101-107.

2. Афанасьев А. А. и др. Помощник инженера для анализа и разработки месторождений нефти и газа на основе интегрированной модели актива // РЯОНЕФТЬ. Профессионально о нефти. - 2024. - Т. 9. - №. 3. - С. 43-49.

3. Печко К. А. и др. Предобработка данных для интегрированного моделирования активов // РЯОНЕФТЬ. Профессионально о нефти. - 2024. -Т. 9. - №. 4. - С. 152-158.

4. Печко К. А. и др. Интерполяция пропусков данных технологического режима скважин алгоритмами машинного обучения // РЯОНЕФТЬ. Профессионально о нефти. - 2023. - Т. 8. - №. 3. - С. 163-166.

5. Бровин Н. М. и др. Совершенствование алгоритма оптимизации графика ввода в эксплуатацию добывающих скважин на интегрированной модели. - 2022.

6. Печко К. А., Сенькин И. С., Белоногов Е. В. Моделирование скважин методами машинного обучения для задач интегрированного моделирования // РЯОНЕФТЬ. Профессионально о нефти. - 2022. - Т. 7. -№. 2. - С. 114-120.

Результаты интеллектуальной деятельности:

1. Патент Евразийский № 048889, 30.01.2025. Афанасьев А. А., Печко К. А., Симонов М. В. СПОСОБ ОПРЕДЕЛЕНИЯ ПРОФИЛЯ ДОБЫЧИ НЕФТИ И ГАЗА // Сертификат:1650024017000.

2. Программа для ЭВМ № 2024682559, 25.09.2024. Афанасьев А. А., Печко К. А., Симонов М. В., Коротеев Д. А., Орлов Д. М., Петросянц М. Т., Илларионов Е. А., Ишимбаев М. В., Волосков Д. С. Программный комплекс метамоделирования актива // Заявка:2024681884.

3. Программа для ЭВМ № 2024665651, 03.07.2024. Афанасьев А. А., Печко К. А., Симонов М. В. Программа для определения перепада давления в скважине // 3аявка:2024665063.

Структура и объем диссертации.

Диссертация состоит из введения, четырех глав, заключения и двух приложений. Полный объем диссертации составляет 244 страницы с 9 рисунками и 12 таблицами. Список литературы содержит 87 наименований.

Содержание работы

Во введении: последовательно раскрывается обоснование актуальности исследования, связанного с необходимостью повышения скорости и точности моделирования процессов разработки нефтегазовых месторождений. Постулируется, что современные условия эксплуатации характеризуются сложностью, гетерогенностью исходных данных и жесткими ограничениями на вычислительные ресурсы. Приводится анализ традиционных методов математического моделирования, основанных на физических законах, отмечается их высокая детализация и одновременно значительные вычислительные издержки, особенно при учете множества взаимосвязанных факторов и необходимости моделирования крупных и комплексных объектов.

В качестве перспективного решения рассматривается многомасштабное моделирование, сочетающее детализированные физические модели с вычислительно эффективными суррогатными аппроксимациями, построенными с использованием методов машинного обучения. Делается акцент на том, что подобная интеграция позволяет существенно ускорить расчеты без потери точности, однако отмечается, что существующие подходы часто сталкиваются с проблемами адаптации суррогатных моделей к изменяющимся условиям эксплуатации, что требует ручной настройки и значительных вычислительных ресурсов для их переобучения.

Отдельно подчеркивается, что внедрение механизмов активного обучения суррогатных моделей открывает новые возможности для повышения эффективности моделирования. Приводится аргументация в пользу автоматизированного отбора наиболее информативных данных для обучения, позволяющего минимизировать объем необходимых высокоточных вычислений и одновременно повысить точность прогнозирования, что особенно востребовано в задачах оперативного управления добычей, где требуется быстрая адаптация моделей к новым данным.

В качестве одной из стратегических задач указывается создание методов многомасштабного моделирования на основе активного обучения суррогатных моделей, что позволит формировать более эффективные цифровые двойники месторождений с высокой скоростью и точностью расчетов.

Определяется объект исследования — процессы моделирования и оптимизации разработки нефтегазовых месторождений с применением методов математического и компьютерного моделирования, а также предмет исследования — методы, алгоритмы и программное обеспечение гибридного моделирования нефтегазовых процессов на базе активного обучения суррогатных моделей. Формулируется цель исследования, заключающаяся в разработке и внедрении методов многомасштабного моделирования нефтегазовых процессов с использованием активного обучения суррогатных моделей, обеспечивающих оперативный анализ большего числа сценариев разработки и повышение обоснованности принимаемых решений в условиях ограниченности времени и вычислительных ресурсов.

В работе приводится перечень задач, последовательное решение которых необходимо для достижения поставленной цели. В их числе — разработка методов и алгоритмов сопряжения моделей различной детальности в рамках интегрированной модели актива, построение суррогатных моделей на базе машинного обучения, создание эффективных процедур активного выбора информативных точек для обучения прокси-моделей и программная реализация предложенных решений с последующей верификацией на реальных и синтетических задачах.

В разделе научной новизны подчеркивается, что впервые разработаны методы многомасштабного моделирования нефтегазовых процессов, основанные на интеграции детальных физических моделей с суррогатными аппроксимациями, строящимися с использованием активного обучения. Приводится описание подхода, обеспечивающего автоматизированный выбор информативных точек для обучения суррогатных моделей в рамках интегрированного моделирования активов, что позволяет существенно снизить вычислительные затраты при сохранении высокой точности моделирования. Указывается на реализацию

программного комплекса, демонстрирующего эффективность предложенных методов на синтетических и реальных тестах.

Теоретическая значимость работы связывается с развитием научных подходов к моделированию сложных инженерных задач в условиях ограниченных ресурсов, а также с расширением фундаментальных основ построения эффективных вычислительных схем для задач моделирования и оптимизации разработки нефтегазовых месторождений, что способствует дальнейшему развитию теории многомасштабных и гибридных вычислительных методов.

Практическая значимость исследования заключается в создании и внедрении эффективных методов и программных средств, которые позволяют существенно снизить вычислительные затраты и повысить оперативность анализа при моделировании и оптимизации разработки нефтегазовых месторождений. Делается акцент на интеграции разработанных алгоритмов в корпоративные цифровые платформы, что способствует повышению точности и скорости инженерных расчетов, обоснованности принимаемых решений и эффективности эксплуатации нефтегазовых активов в реальных производственных задачах.

В разделе, посвященном методам исследования, перечисляются примененные подходы — математическое и компьютерное моделирование, машинное обучение, методы оптимизации, математическая статистика и программная инженерия.

Приводятся ключевые положения, выносимые на защиту, в числе которых — методология и алгоритмы суррогатного моделирования нефтегазовых процессов на основе машинного обучения, методы гибридного многомасштабного моделирования с использованием активного обучения для оптимального распределения вычислительных ресурсов, а также программная реализация указанных методов для автоматизированного построения и сопряжения моделей различной детальности с их верификацией на синтетических и промышленных задачах.

Достоверность научных положений, выводов и практических рекомендаций обосновывается эмпирическим подходом при разработке методов и алгоритмов, верификацией на синтетических и реальных данных, сопоставлением с

результатами численного моделирования и промысловыми данными, успешной апробацией в производственных условиях, сравнением с альтернативными методами и независимой экспертизой в научном сообществе через публикации, доклады и патенты.

В разделе апробации отмечается, что результаты исследования были представлены на ведущих научных конференциях по тематике цифровизации и интеллектуального анализа данных в нефтегазовой отрасли, а также прошли опытно-промышленную проверку на активах компании «Газпромнефть».

В заключительной части введения приводится информация о личном вкладе автора, в который входит участие во всех этапах разработки, внедрения и валидации методов, программных решений и публикационной активности, а также сведения о публикациях и результатах интеллектуальной деятельности, подтверждающих значимость и оригинальность выполненного исследования.

В первой главе:

диссертационной работы проводится систематизированный и детализированный обзор методов моделирования нефтегазовых процессов, акцентируя внимание на современных тенденциях цифровизации и возрастающих требованиях к эффективности и точности инженерных расчетов. В работе подчеркивается, что цифровая трансформация нефтегазовой отрасли обусловлена усложнением геологической среды, ростом неопределенностей и ужесточением экономических требований к разработке месторождений. В качестве центрального направления современной практики позиционируется интеграция интегрированных моделей актива (ИМА), которые охватывают процессы в пласте, скважинах и наземной инфраструктуре, реализуя принципы цифрового двойника месторождения.

Похожие диссертационные работы по специальности «Другие cпециальности», 00.00.00 шифр ВАК

Список литературы диссертационного исследования кандидат наук Печко Константин Анатольевич, 2025 год

Список литературы

1. Forrester, A. I. J. Multi-fidelity optimization via surrogate modelling /

A. I. J. Forrester, A. Sobester, A. J. Keane // Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences. — 2007. — Vol. 463, no. 2088. —P. 3251—3269. —URL: https://royalsocietypublishing.org/doi/ 10.1098/rspa.2007.1900.

2. Multifidelity Dimension Reduction via Active Subspaces / R. Lam [et al.] // SIAM/ASA Journal on Uncertainty Quantification. — 2020. — Vol. 8, no. 2. — P. 572—601. — URL: https://doi.org/10.1137/19M1261427.

3. Сенькин, И. С. Оптимизация расчетов интегрированных моделей [Электронный ресурс] / И. С. Сенькин, Л. И. Акмадиева, Е. В. Белоногов // PROнефть. — 2020. — Т. 7, № 2. — С. 41—51. — URL: https://proneft.elpub. ru/jour/article/view/252 (дата обр. 12.06.2025).

4. Application of Machine Learning in Integrated Modeling of the Oil and Gas Fields / K. Pechko [et al.]. — 2023. — URL: https://doi.org/10.3997/2214-4609.202332061.

5. Reisenthel, P. Statistical benchmarking of surrogate-based and other optimization methods constrained by fixed computational budget / P. Reisenthel, D. Lesieutre // 18th AIAA/ASME/AHS Adaptive Structures Conference. — AIAA, 2010. — P. 30—88.

6. Razavi, S. Review of surrogate modeling in water resources / S. Razavi,

B. A. Tolson, D. H. Burn // Water Resources Research. — 2012. — Т. 48, № 7. — W07401.

7. Palagi, S. Integrated asset modeling for oil and gas production optimization / S. Palagi, L. J. Durlofsky // Journal of Petroleum Science and Engineering. — 2016. — Vol. 147. — P. 550—566.

8. Aziz, K. Integrated Asset Modeling: A New Frontier in Reservoir Simulation / K. Aziz, A. Settari // SPE Journal. — 2010. — Vol. 15, no. 2. — P. 305—313.

9. Наумович, А. В. Применение суррогатного моделирования в задачах нефтегазовой промышленности / А. В. Наумович, И. Ю. Михайлова // Нефтегазовое дело. — 2022. — Т. 20, № 3. — С. 55—68.

10. Heidari, P. Digital twin modeling in upstream oil and gas industry: Opportunities and challenges / P. Heidari, M. Ghassemi, S. Yousefi // Journal of Natural Gas Science and Engineering. — 2021. — Vol. 95. — P. 104225.

11. Patel, R. Active Learning for Efficient Surrogate Modeling in Oil Reservoir Simulation / R. Patel, R. Misener // Proceedings of the European Symposium on Computer Aided Process Engineering. — 2015. — P. 203—208.

12. Wang, J. Surrogate Modeling and Machine Learning in Reservoir Engineering: A Review / J. Wang, S. Sun // Computational Geosciences. — 2021. — Vol. 25. — P. 1301—1332.

13. Society of Petroleum Engineers (SPE). Reservoir Simulation / Society of Petroleum Engineers (SPE). — 2025. — URL: https : / / petrowiki . spe . org/Reservoir _ simulation ; Online encyclopedia entry, accessed June 2025. PetroWiki, Society of Petroleum Engineers.

14. Coats, K. H. Compositional and Black Oil Reservoir Simulation / K. H. Coats, L. K. Thomas, R. G. Pierson // SPE Reservoir Evaluation & Engineering. — 1998. — Vol. 1, no. 04. — P. 372—379.

15. Parallel Multi-Stage Preconditioners with Adaptive Setup for the Black Oil Model / L. Zhao [et al.] // arXiv preprint. — 2022. — Vol. arXiv:2201.01970. — URL: https : / / arxiv . org / abs / 2201 . 01970 ; Preprint, submitted to Computers and Geosciences.

16. N'diaye, M. Comparison of Nonlinear Field-Split Preconditioners for Two-Phase Flow in Heterogeneous Porous Media / M. N'diaye, F. P. Ha-mon, H. A. Tchelepi // arXiv preprint. — 2022. — May. — Vol. arXiv:2205.05913. — URL: https : / / arxiv . org / abs / 2205 . 05913 ; Preprint.

17. Speeding up the reservoir simulation by real-time prediction of the initial guess for the Newton-Raphson's iterations / M. Petrosyants [et al.] // Computational Geosciences. — 2024. — Vol. 28, no. 4. — P. 605—613.

18. Parallel Multi-Stage Preconditioners with Adaptive Setup for the Black Oil Model / L. Zhao [et al.] // Computers & Geosciences. — 2022. — Vol. 168. — P. 105230. — Published online Sep 2022.

19. Efficient reordered nonlinear Gauss-Seidel solvers with higher order for black-oil models / O. S. Klemetsdal [et al.] // Computational Geosciences. — 2020. — Vol. 24, no. 2. — P. 593—607.

20. Sheth, S. Inexact Newton Method for General Purpose Reservoir Simulation / S. Sheth, A. Moncorge // arXiv preprint. — 2019. — Т. arXiv:1912.06568. — URL: https://arxiv.org/abs/1912.06568 ; Applies adaptive forcing-term strategies to reduce oversolving and total linear iterations in fully-coupled three-phase (including gas injection) reservoir models.

21. Petroleum Reservoir Simulation: The Engineering Approach [Электронный ресурс] / L. W. Lake [et al.]. — Society of Petroleum Engineers, 2014. — P. 516. —URL: https://www.geokniga.org/bookfiles/geokniga-petroleum-reservoir-simulation.pdf (visited on 06/12/2025).

22. Ahmed, T. H. Reservoir Engineering Handbook / T. H. Ahmed. — 4th ed. — Boston, MA : Gulf Professional Publishing, 2010.

23. Dake, L. P. Fundamentals of Reservoir Engineering / L. P. Dake. — 1st ed. — Amsterdam, The Netherlands : Elsevier Science, 1983. — (Developments in Petroleum Science, vol. 8).

24. Sakhaei, Z. Assessment of Empirical/Theoretical Relative Permeability Correlations for Gas-Oil/Condensate Systems / Z. Sakhaei, R. Azin, S. Osfouri // Proceedings of the 1st Biennial Persian Gulf Oil, Gas and Petrochemical Energy and Environment Conference. Vol. 20. — Bushehr, Iran, 04/2016. — P. 1—20.

25. Corey, A. T. The Interrelation Between Gas and Oil Relative Permeabilities / A. T. Corey // Producers Monthly. — 1954. — Vol. 19, no. 1. — P. 38—41.

26. Standing, M. B. A Pressure-Volume-Temperature Correlation for Mixtures of California Oils and Gases / M. B. Standing // Drilling and Production Practice. — 1947. — P. 275—287.

27. Vasquez, M. E. Correlations for Fluid Physical Property Prediction / M. E. Vasquez, H. D. Beggs // Journal of Petroleum Technology. — 1980. — Vol. 32, no. 6. — P. 968—970.

28. Aziz, K. Petroleum Reservoir Simulation / K. Aziz, A. Settari. — 1st ed. — London, UK : Applied Science Publishers, 1979. — P. 476.

29. Bear, J. Dynamics of Fluids in Porous Media / J. Bear. — Reprint of 1972 edition. — New York, USA : Dover Publications, 1972.

30. Применение интегрированного моделирования в нефтегазовой отрасли / Е. В. Филиппов [и др.] // Вестник ПНИПУ. Геология. Нефтегазовое и горное дело. — 2020. — Т. 20, № 4. — С. 386—400. — URL: https://ered.pstu. ru/index.php/geo/article/view/1367 (дата обр. 12.06.2025) ; Электронный ресурс.

31. Data-driven methods for flow and transport in porous media: a review / G. Yang [et al.] // International Journal of Heat and Mass Transfer. — 2024. — Vol.235. — P. 126149. — URL: https ://www. sciencedirect. com/science/article/pii/S0017931024005652 (visited on 06/20/2025).

32. An integrated framework for accelerating reactive flow simulation using GPU and machine learning models / R. Mao [et al.] // Computational Physics Communications. — 2024. — Vol. XXX. — YYY—ZZZ. — URL: https: //www.sciencedirect.com/science/article/pii/S1540748924003201 (visited on 06/20/2025).

33. DeepFlame: A deep learning empowered open-source platform for reacting flow simulations / R. Mao [et al.] // Proceedings of the Combustion Institute. — 2022. — Vol. XX. —P. 112—119. —URL: https://arxiv.org/abs/2210.07094 (visited on 06/20/2025).

34. Jiang, J. A Multigrid Graph U-Net Framework for Simulating Multiphase Flow in Heterogeneous Porous Media / J. Jiang, J. Chen, Z. Yang // arXiv preprint. — 2024. — Vol. arXiv:2412.12757. — URL: https://arxiv.org/abs/ 2412.12757 (visited on 06/20/2025).

35. Model-based decision analysis applied to petroleum field development and management [Электронный ресурс] / D. J. Schiozer [et al.] // Oil & Gas Science and Technology - Revue d'IFP Energies nouvelles. — 2019. — Vol. 74. — P. 46. —URL: https://ogst.ifpenergiesnouvelles.fr/articles/ogst/full%5C_ html/2019/01/ogst19019/ogst19019.html (visited on 06/12/2025).

36. The Benefits of Integrated Asset Modeling: Lessons Learned from Field Cases [Электронный ресурс] / M. Rotondi [et al.] // SPE Europec/EAGE Annual Conference and Exhibition. — 2008. — URL: https : / / onepetro . org / SPEEUROPEC / proceedings / 08EUR0 / All - 08EUR0 / SPE - 113831 -MS/120711 (visited on 06/12/2025).

37. Digital Twin: Enabling Technologies, Challenges and Open Research / A. C. Bittencourt [et al.] // Computers & Chemical Engineering. — 2018. — Vol.112. —P. 25—45. —URL: https://www.sciencedirect.com/science/ article/pii/S0098135418302569 (visited on 06/12/2025).

38. Robust optimization of production strategies under geological uncertainty using multi-objective ensemble-based optimization / R. Bordas [et al.] // Computational Geosciences. — 2020. — Vol. 24, no. 4. — P. 1011—1030. — URL: https://link.springer.com/article/10. 1007/s10596-020-09943-8 (visited on 06/12/2025).

39. Dehdari, V. Comparison of optimization algorithms for reservoir management with constraints —a case study / V. Dehdari, D. S. Oliver, C. V. Deutsch // Journal of Petroleum Science and Engineering. — 2012. — Vol. 100. — P. 41—49. — URL: https://doi.org/10.1016/j.petrol.2012.11.013 (visited on 06/12/2025).

40. Kor, P. Reservoir Production Management With Bayesian Optimization: Achieving Robust Results in a Fraction of the Time / P. Kor, A. Hong, R. B. Bratvold // SPE Journal. — 2023. — Vol. 29, no. 02. — P. 620—640.

41. Kudela, J. Combining Lipschitz and RBF Surrogate Models for High-dimensional Computationally Expensive Problems / J. Kudela, R. Matousek // Information Sciences. — 2023. — Vol. 619. — P. 457—477.

42. Surrogate-assisted Evolutionary Algorithms for Expensive Combinatorial Optimization: A Survey / S. Liu [et al.] // Complex & Intelligent Systems. — 2024. — Vol. 10, no. 4. — P. 5933—5949.

43. Pierce, C. M. A Comparison-Relationship-Surrogate Evolutionary Algorithm for Multi-Objective Optimization / C. M. Pierce, Y.-K. Kim, I. Bazarov // Swarm and Evolutionary Computation. — 2025. — Vol. 95. — P. 101947.

44. Surrogate-assisted Global and Distributed Local Collaborative Optimization Algorithm for Expensive Constrained Optimization Problems / X. Liu [et al.] // Scientific Reports. — 2025. — Vol. 15. — P. 1728.

45. Shahriari, B. Unbounded Bayesian Optimization via Regularization / B. Shahriari, A. Bouchard-Côté, N. de Freitas // Proceedings of the 19th International Conference on Artificial Intelligence and Statistics (AISTATS). Vol. 51. — Cadiz, Spain : PMLR, 05/2016. — P. 1168—1176.

46. Jin, Y. Surrogate-assisted Evolutionary Computation: Recent Advances and Future Challenges / Y. Jin // Swarm and Evolutionary Computation. — 2011. — Vol. 1, no. 2. — P. 61—70.

47. Forrester, A. I. J. Engineering Design via Surrogate Modelling: A Practical Guide / A. I. J. Forrester, A. Sóbester, A. J. Keane. — 1st ed. — Chichester, West Sussex, UK : John Wiley & Sons, 2008.

48. Surrogate-Based Modeling and Optimization: Applications in Engineering / ed. by S. Koziel, L. Leifsson. — 1st ed. — New York, NY, USA : Springer, 2013.

49. Jiang, S. Deep-learning-based surrogate model for reservoir simulation with time-varying well controls / S. Jiang, Y. Jiang, L. J. Durlofsky // Journal of Petroleum Science and Engineering. — 2020. — Sept. — Vol. 192. — P. 107273. — Available online 22 May 2020.

50. Lataniotis, C. Extending Classical Surrogate Modelling to High-Dimensions through Supervised Dimensionality Reduction: A Data-Driven Approach / C. Lataniotis, S. Marelli, B. Sudret // arXiv preprint. — 2018. — Vol. arXiv:1812.06309. — URL: https://arxiv.org/abs/1812.06309.

51. Machine-Learned Surrogate Models for Efficient Oil Well Placement Under Operational Reservoir Constraints / S. M. Mousavi [et al.] // SPE Journal. — 2023. — Vol. 29, no. 1. — P. 518—537.

52. Deep Graph Learning-Based Surrogate Model for Inverse Modeling of Fractured Reservoirs / X. Ma [et al.] // Mathematics. — 2024. — Vol. 12, no. 5. — P. 754.

53. Tang, H. Graph Network Surrogate Model for Subsurface Flow Optimization / H. Tang, L. J. Durlofsky // Journal of Computational Physics. — 2024. — Vol. 512. — P. 113132.

54. Settles, B. Active Learning Literature Survey / B. Settles // University of Wisconsin-Madison, Computer Sciences Technical Report. — 2009. — No. 1648.

55. Active Learning: A Survey / C. C. Aggarwal [et al.] // Data Classification: Algorithms and Applications. — 2014. — P. 571—605.

56. Seung, H. S. Query by Committee / H. S. Seung, M. Opper, H. Sompolinsky // Proceedings of the Fifth Annual ACM Workshop on Computational Learning Theory. — 1992. — July. — Vol. 3, no. 1. — P. 287—294. — URL: https://doi.org/10.1145/130385.130417 ; Conference held July 27-29, 1992, Pittsburgh, PA, USA.

57. Toward Optimal Active Learning through Sampling Estimation of Error Reduction. Vol. 7 / ed. by P. Kidwelly. — ICML. Morgan Kaufmann, 07/2001. — (7). — ISBN: 1-55860-778-1.

58. Queries and Concept Learning. Vol. 24 / ed. by D. ANGLUIN. — Machine Learning. Kluwer Academic Publishers, 07/1998. — (7). — ISBN: 1-55860-778-1.

59. Cacciarelli, D. Active Learning for Data Streams: A Survey / D. Cacciarelli, M. Kulahci // Machine Learning. — 2024. — Vol. 113. — P. 185—239.

60. Zhu, J.-J. Generative Adversarial Active Learning / J.-J. Zhu, J. Bento // arXiv preprint. — 2017. — Vol. arXiv:1702.07956. — URL: https://arxiv. org/abs/1702.07956.

61. Semantic-aligned Query Synthesis for Active Learning / Y.-P. Tang [et al.] // ICLR 2025 Workshop on MLGenX. — 2025. — URL: https://openreview. net/forum?id=CZvbXXgjrn ; "Generative Membership Query Descriptor (Gen-MQD)" for AL.

62. Moroshko, E. Selective Sampling with Drift / E. Moroshko, K. Crammer // Proceedings of the Seventeenth International Conference on Artificial Intelligence and Statistics (AISTATS). Vol. 33. — Reykjavik, Iceland : PMLR, 04/2014. — P. 651—659.

63. Streaming Active Learning with Deep Neural Networks / A. Saran [et al.] // Proceedings of the 40th International Conference on Machine Learning (ICML). Vol. 202. — Vienna, Austria : PMLR, 07/2023. — P. 30005—30021.

64. Dekel, O. Selective Sampling and Active Learning from Single and Multiple Teachers / O. Dekel, C. Gentile, K. Sridharan // Journal of Machine Learning Research. — 2012. — Vol. 13. — P. 2655—2697.

65. Dekel, O. Selective Sampling and Active Learning from Single and Multiple Teachers / O. Dekel, C. Gentile, K. Sridharan // Proceedings of the 23rd International Conference on Learning Theory (COLT). — OML, 2010. — P. 346—358.

66. Adaptive Diffusion in Graph Neural Networks / J. Zhao [et al.] // Advances in Neural Information Processing Systems. — 2021. — Dec. — Vol. 34, no. 2. — P. 1—34. — NeurIPS 2021.

67. Zhu, J. Improving ML-Based Information Retrieval Software with User-Driven Functional Testing and Defect Class Analysis / J. Zhu, T. Long, W. Wang // Proceedings of the 30th ACM Joint European Software Engineering Conference and Symposium on the Foundations of Software Engineering. — 2022. — Nov. — No. 1. — P. 1—12. — ESEC/FSE '22: November 14-18, 2022, Singapore.

68. Yu, J. Active Learning for Well Control Optimization with Surrogate Models / J. Yu, B. Jafarpour // SPE Journal. — 2022. — Oct. — Vol. 27, no. 5. — P. 2668—2688. — Available to Purchase.

69. An active learning surrogate-assisted multi-objective optimization algorithm / L. Wang [et al.] // Engineering Applications of Artificial Intelligence. — 2025. — in press.

70. Nonlinear information fusion algorithms for data-efficient multi-fidelity modelling / P. Perdikaris [h gp.] // Proceedings of the Royal Society A. — 2017. — T. 473, № 2198. — C. 20160751.

71. Raissi, M. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations / M. Raissi, P. Perdikaris, G. E. Karniadakis // Journal of Computational Physics. — 2019. — T. 378. — C. 686—707.

72. Kennedy, M. C. Predicting the Output from a Complex Computer Code When Fast Approximations Are Available / M. C. Kennedy, A. O'Hagan // Biometrika. — 2000. — Vol. 87, no. 1. — P. 1—13.

73. Nguyen, T. H. Deep Multi-Fidelity Modeling for Engineering Systems / T. H. Nguyen, D. D. Tuyen, D. H. Minh // Mathematics. — 2023. — T. 11, № 4. — C. 877.

74. Frazier, P. I. A Tutorial on Bayesian Optimization / P. I. Frazier // arXiv preprint arXiv:1807.02811. — 2018.

75. Buckley, S. E. Mechanism of Fluid Displacements in Sands / S. E. Buckley, M. C. Leverett // Transactions of the AIME. — 1942. — Vol. 146, no. 1. — P. 107—116.

76. Aziz, K. Petroleum Reservoir Simulation / K. Aziz, A. Settari. — London : Applied Science Publishers, 1979. — P. 497.

77. Fetkovich, M. J. A Simple Approach to Water Influx Calculations-Finite Aquifer Systems /M.J. Fetkovich // Journal of Petroleum Technology. — 1971. — Vol. 23, no. 7. — P. 814—828.

78. White, F. M. Fluid Mechanics / F. M. White // McGraw—Hill Education. — 2016. — P. 345—378.

79. Advances and outlook of integrated reservoir-wellbore-pipe network simulation technology [Электронный ресурс] / H. Yu [et al.] // Energy Geoscience. —

2024. — Vol. 5, no. 3. — P. 100315. — URL: https://www.sciencedirect. com/science/article/pii/S2666759224000301 (visited on 06/13/2025).

80. Services, S. O. Defining Integrated Asset Modeling [Электронный ресурс] / S. O. Services // Oilfield Review. — 2016. — May. — URL: https://www. slb.com / resource- library / article / 2016 / defining- integrated-asset-modeling (visited on 06/13/2025).

81. Thenon, A. Multi-fidelity Proxy Models for Reservoir Engineering / A. Thenon, V. Gervais, M. Le Ravalec // Proceedings of the ECMOR XV - 15th European Conference on the Mathematics of Oil Recovery. — 08.2016. — URL: https: / / www. researchgate. net / publication / 307573448 _ Multi- fidelity _ Proxy _ Models_for_Reservoir_Engineering.

82. Simulation and optimization of a coupled reservoir and multi-phase flow network model / P. R. Leinan [и др.] // Advances in Geo-Energy Research. —

2025. — Т. 15, № 3. — С. 203—215. — URL: https://doi.org/10.46690/ager. 2025.03.04.

83. Conti, S. Multi-fidelity optimization via surrogate modelling / S. Conti, S. Marelli, B. Sudret // Computational Methods in Applied Mechanics and Engineering. — 2024. — in press.

84. Multi-fidelity co-kriging surrogate model for ship hull form optimization / Q. Zhou [et al.] // Engineering Applications of Computational Fluid Mechanics. — 2021. — Vol. 15, no. 1. — P. 795—814.

85. Multi-fidelity transfer stacking surrogate model for multi-objective optimization / K. Xu [et al.] // Expert Systems with Applications. — 2023. — Vol. 223. — P. 119965.

86. Интерполяция пропусков данных технологического режима скважин алгоритмами машинного обучения / К. А. Печко [и др.] // РИОНЕФТЬ. Профессионально о нефти. — 2023. — Т. 8, № 3. — С. 163—166. — URL: https://proneft.elpub.ru/jour/article/view/448.

87. Bahrami, P. A Review of Proxy Modeling Highlighting Applications for Reservoir Engineering / P. Bahrami, F. Sahari Moghaddam, L. A. James // Energies. — 2022. — Vol. 15, no. 14. — P. 5247. — Open access under CC BY 4.0.

177

Приложение l

ЕА3 00/ 20 17

О

о

о_

н

^

ш

о

На № 522НТЦ/208ОРД от 19/08/2024 ]

Номер заявки: 202490130 Дата отправки 19.11.2024

Российская Федерация, г. Санкт-Петербург, ул. Почтамтская, д.3-5, ПАО "Газпром нефть", Дирекция по технологическому развитию евразийскому патентному поверенному г-же Чугункиной Л.А.

УВЕДОМЛЕНИЕ

о готовности выдать евразийский патент

В результате рассмотрения данной заявки коллегией экспертов ЕАПВ установлены соответствие заявленного изобретения (группы изобретений) условиям патентоспособности, предусмотренным правилом 3 Патентной инструкции к Евразийской патентной конвенции (далее - Инструкция), и возможность выдачи евразийского патента с: формулой изобретения Пв первоначальной редакции й в редакции от 19/08/2024 описанием изобретения Пв первоначальной редакции И в редакции от 19/08/2024 рефератом Пв первоначальной редакции Н в редакции от 19/08/2024

чертежами ЕЗв первоначальной редакции □ в редакции от

1. Решение о выдаче евразийского патента будет принято при условии:

Н уплаты заявителем установленных пошлин, в том числе: _И пошлины за выдачу евразийского патента в размере 27000 * рублей РФ_

Пошлина за выдачу евразийского патента должна быть уплачена в установленном размере в течение 4-хмесяцев с даты направления данного уведомления (правило 47(3) Инструкции, пункты 1(5), 5(1) Положения о пошлинах Евразийской патентной организации (далее-ЕАПО)). Если в течение этого срока заявитель не уплатит пошлину за выдачу евразийского патента, он может уплатить ее в течение 2-х месяцев с дат ы окончания вышеуказанного четырехмесячного срока при условии уплаты дополнительной пошлины в размере 4500 рублей РФ (правило 47(3) Инструкции, пункты 1(5), 5(3) Положения о пошлинах ЕАПО).

□ дополнительной пошлины за публикацию материалов евразийского патента в размере

рублей РФ

(количество листов в заявке: 28, включая формулу изобретения, описание изобретения, реферат, чертежи и иные подлежащие публикации материалы)

Согласно правилу 50(4) Инструкции дополнительная пошлина уплачивается за публикацию евразийского патента, если в подлежащих публикации материалах содержится более 35 листов.

Дополнительная пошлина должна быть уплачена в установленном размере в течение 6-ти месяцев (правило 50(4) Инструкции, пункты 1(5), 5(2) Положения о пошлинах ЕАПО) с даты направления данного уведомления.

В случае неуплаты указанных пошлин выдача и публикация евразийского патента не производятся, а заявка считается отозванной.

2.П Заявителю предлагается внести следующие изменения в

□ формулу изобретения, пункт ы

□ описание изобретения, страницы

□ чертежи, номера

□ реферат

Копии страниц П прилагаются П не прилагаются Заявителю следует представить один экземпляр^ формулы, □ реферата, □ чертежей в течение срока, указанного в п.1 настоящего уведомления.

3. Название изобретения при публикации описания изобретения будет приведено: □ в первоначальной редакции:

□ по заявлению о выдаче евразийского патента

□ по описанию к изобретению

= [

И в редакции, изложенной в корреспонденции от 19/08/2024 □ в предлагаемой редакции ()

Главный эксперт

Отдела механики, физики и

электротехники

ДОИУИРМТ Г-ПгП.ПИГ-1Ч ......... ИОилниА

вчдагц- СМ=Зизтдинов Н.Е. Лн*.*.. ^.чк 20.07.2023-13.07.2023

Н.В. Зиатдинов

+7(495)411-61-61*324

Указан размер пошлин, действующий на момент отправки данного уведомления. При оплате заявитель должен руководствоваться размером пошлин, действующим на дату их уплаты.

о

о

о_

н

^

ш

о

ПРОТОКОЛ заседания коллегии экспертов Евразийского патентного ведомства Евразийской патентной организации

18 ноября 2024 г.

Евразийская заявка № 202490130/31 Присутствовали:

Члены коллегии:

Н.В. Зиатдинов С.С. Нурушева Д.Ф. Крылов

-председатель

Заявитель:

(представитель заявителя)

Изобретатель(и):

Предмет рассмотрения: Предметом рассмотрения является заявка на выдачу евразийского патента № 202490130/31 Заявка рассмотрена в полном объеме.

Результат рассмотрения: В результате рассмотрения всех материалов заявки с учетом источников информации, указанных в отчете о патентном поиске, обстоятельств, препятствующих выдаче евразийского патента по данной заявке, не установлено.

Мнение заявителя (представителя заявителя):

= « Мнение изобретателя(ей):

= а Особое мнение экспертов:

Подписи членов коллегии:

Дануч<гит гарпигач ^пмрйнмин ИйщщПые!

1689860440481 ицдаг* С^Зиэтдинов Н.Е. лш::г., ^.гчк 20.07.2023-18.07.2028

Н.В. Зиатдинов

1 Данучгнт г-пгп--1иг-1ч -М1*-Н ^рйим мы ПВД^МСьШ

1689860416289

1 СМ=Нуру!иева С.С.

до 1 гыи-.гик 20.07.2023-18.07.2028

С.С. Нурушева

Дануч<гит гчрпигач

^пмрйнмин |1^и.Л><1С|(1

1689860504361 ицдаг* С ^Крылов Д.(Ь. Лг«:: ► - ч'-чн 20.07.2023-18.07.2028

Д.Ф. Крылов

РОССИЙСКАЯ ФЕДЕРАЦИЯ

RU2024682559

ФЕДЕРАЛЬНАЯ СЛУЖБА ПО ИНТЕЛЛЕКТУАЛЬНОЙ СОБСТВЕННОСТИ

ГОСУДАРСТВЕННАЯ РЕГИСТРАЦИЯ ПРОГРАММЫ ДЛЯ ЭВМ

Номер регистрации (свидетельства): Автор(ы):

2024682559 Афанасьев Александр Андриянович ^Ц),

Дата регистрации: 25.09.2024 Симонов Максим Владимирович ^Ц),

Номер и дата поступления заявки: Печко Константин Анатольевич

2024681884 25.09.2024 Коротеев Дмитрий Анатольевич

Дата публикации и номер бюллетеня: Орлов Денис Михайлович ^Ц),

25.09.2024 Бюл. № 10 Петросянц Мушег Тигранович

Илларионов Егор Александрович

Ишимбаев Марсель Венерович ^Ц),

Волосков Дмитрий Сергеевич ^Ц)

Правообладатель(и):

Публичное акционерное общество «Газпром

нефть» ^и)

Название программы для ЭВМ:

Программный комплекс метамоделирования актива

Реферат:

Программа для ЭВМ предназначена для совместных расчетов метамоделей-компонент в интегрированной модели активов. Областью применения программы является интегрированное моделирование активов. Функциональные возможности программы: проведение многовариативных вычислений с высокой скоростью расчета без привлечения больших вычислительных мощностей; учет расписания ввода новой скважины и смены режимов работы скважин; перевод скважин в режим поддержания пластового давления; прогнозирование профиля добычи с учетом ограничений, накладываемых каждым компонентом (моделью).

Язык программирования: Python

Объем программы для ЭВМ: 649,4 Мб

РОССИЙСКАЯ ФЕДЕРАЦИЯ

RU2024665651

ФЕДЕРАЛЬНАЯ СЛУЖБА ПО ИНТЕЛЛЕКТУАЛЬНОЙ СОБСТВЕННОСТИ

ГОСУДАРСТВЕННАЯ РЕГИСТРАЦИЯ ПРОГРАММЫ ДЛЯ ЭВМ

Номер регистрации (свидетельства): Автор(ы):

2024665651 Афанасьев Александр Андриянович ^Ц),

Дата регистрации: 03.07.2024 Симонов Максим Владимирович ^Ц),

Номер и дата поступления заявки: Печко Константин Анатольевич ^Ц)

2024665063 03.07.2024 Правообладатель(и):

Дата публикации и номер бюллетеня: Публичное акционерное общество «Газпром

03.07.2024 Бюл. № 7 нефть» ^Ц)

Название программы для ЭВМ:

Программа для определения перепада давления в скважине Реферат:

Программа предназначена для создания и адаптации модели машинного обучения на исторические данные (устьевое и забойное давления, дебит жидкости, обводненность, газовый фактор, частота работы насоса в случае эксплуатации с электроцентробежным насосом), которые получают с месторождения нефти и газа или с результатов расчета физико-математической модели. Областью применения программы является интегрированное моделирование активов. Функциональные возможности программы: обучение модели на основе входных данных, описывающих состояние скважины; расчет прогнозного забойного давления с помощью созданной модели при различных значениях параметров технологического режима работы скважины.

Язык программирования: Python

Объем программы для ЭВМ: 1,76 Гб

182

Тексты публикаций

253 (2025) 213946

Contents lists available at ScienceDirect

Geoenergy Science and Engineering

journal homepage: www.sciencedirect.com/journal/geoenergy-science-and-engineering

Geoenergy

Science and Engineering

Production rate forecasting using pressure and saturation estimates near the wellbore

D. Voloskova '*, E. Gladchenkoa, D. Akhmetova , E. Illarionov a'b , K. Pechkoc , A. Afanasevd, M. Simonovd, D. Orlova , D. Koroteeva

a Center for Petroleum Science and Engineering, Moscow, Russian Federation b Moscow State University, Moscow, Russian Federation c ITMO University, Saint Petersburg, Russian Federation d Independent Researcher

%

Check for updates

A R T I C L E I N F O

A B S T R A C T

Keywords:

Reservoir simulation Proxy-model Production forecast Machine learning

Production flow rate forecasting is crucial for effective reservoir management decisions. Usually, this forecasting relies on either full-scale reservoir models, which demand significant computational resources, or highly simplified surrogate models like the Capacitance-Resistance Model (CRM). While CRM offers computational efficiency, it faces considerable challenges when applied to multilayer reservoirs or scenarios with varying fluid saturations. We present a novel method for flow rate forecasting, occupying an intermediate position between these two approaches. The model consists of several components: separate machine learning models are used to predict pressure and fluid saturation in each grid cell perforated by the well. The predicted pressure and saturation values are then used to estimate well production rates. This workflow preserves the physical interpretability: outputs of intermediate models represent primary dynamic variables and the weighting coefficients of the flow rate estimation model can be interpreted as connection factors of the grid blocks. At the same time, it overcomes some of the limitations of both full-scale reservoir model and CRM-like surrogate models. Unlike CRM, each well is considered as a collection of cells for better representation of structural heterogeneity. Additionally, by considering fluid saturation as a dynamic property, it can manage problems with varying fluid phase composition. The focus only on values near the wellbore facilitates rapid forecasting compared to full-scale modeling. We test the proposed approach using two reservoir models: a simplistic synthetic reservoir and a model of a real-world reservoir, and compare it with the results obtained using a commercial reservoir simulator and CRM. We observe a better accuracy against the CRM and a reasonable approximation of full-scale simulation results.

1. Introduction

Simulation of multiphase flows in porous media is an essential operational process in various fields of reservoir engineering. Traditionally, such simulations are based on the numerical solution of a system of differential equations describing the dynamics of the reservoir state (pressure and fluid saturation field) in time (Aziz and Settari, 1979). This solution relies on numerical techniques such as finite difference, finite volume, or finite element methods (Chen et al., 2006; Monteagudo and Firoozabadi, 2004). Reservoir simulators based on numerical solutions of Partial Differential Equations (PDEs) can accurately capture complex rock and fluid physics and incorporate complex boundary conditions such as aquifers, complex well geometry, or various well completions. However, simulating real-world reservoirs often requires significant computational resources, complicating the use of these

models to solve many engineering problems, involving multiple simulations, such as reservoir development optimization, history matching, and integrated modeling.

One way to reduce the computational cost of modeling is to use reservoir proxy models. There are several classes of such models: probably the most widely used is the class of Capacitance-Resistance Models (CRMs). These methods are based on electrical engineering analogy and characterize interwell and nearwell media by several integral parameters (Sayarpour et al., 2009), in contrast to a full-scale model where parameters are assigned to each grid cell. It provides a fast prediction of total well fluid production using the least amount of data possible. The required training data set includes only historical production and injection rates and producers' bottomhole pressure (BHP) values. Changes in

* Corresponding author.

E-mail address: voloskovdmitriy@gmail.com (D. Voloskov).

https://doi.org/10.10167j.geoen.2025.213946

Received 20 May 2024; Received in revised form 5 February 2025; Accepted 26 April 2025 Available online 13 May 2025

2949-8910/© 2025 Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.

reservoir pressure and geological properties are implicitly accounted for: geological properties are embedded in characteristic coefficients determined during the history matching process.

Considering CRMP, which is one of the many implementations of CRM (Holanda et al., 2018), characteristic coefficients reflect the connectivity between each producer-injector pair and are strongly related to porosity, permeability, etc. On the other hand, they are a limitation of the model: the set of identified characteristic parameters is assumed to be constant, while in real life reservoir state and technological parameters of the wells change with time affecting interwell connectivity (Altaheini et al., 2016). Another limitation of the model is that additional fractional flow models must be tuned to predict different phases (oil and water) separately. One of the most common approaches is to use the Gentil, or Power-Law, model (Gentil, 2005), as it stays within the CRM philosophy and only requires production data to perform the history matching. However, it should be noted that this empirical relationship is not suitable for cases with dramatic changes in saturation. Thus, such cases as water breakthroughs cannot be correctly predicted by the model (Cao et al., 2015).

Another group of methods is based on the estimation of reservoir state dynamics in a lower dimensional domain, usually defined by the proper orthogonal decomposition (POD) basis (see, for example, Voloskov and Pissarenko, 2021; Cardoso and Durlofsky, 2010). In these methods, a lower-dimensional representation of the state variables is estimated either by solving equations in the reduced order domain using the Galerkin or Petrov-Galerkin projection (Carlberg et al., 2017), or by using data-based reduced order time-dynamics approximations (Cardoso and Durlofsky, 2010; Trehan and Durlofsky, 2016).

One more class of proxy modeling methods that has been actively developed in recent times is machine learning methods (for a review see, for example, Samnioti and Gaganis, 2023a,b). Kani and Elsheikh (2018) used a recurrent neural network to model two-phase flow in the POD domain. A group of methods uses variational autoencoders (VAEs) to obtain a representation of the reservoir state in reduced-order subspace (Jin et al., 2020; Lee and Carlberg, 2019; Temirchev et al., 2020; Illarionov et al., 2022). However, despite the attractive power of machine learning methods, these techniques may lack information about the internal structure of complex processes, resulting in challenges when generalizing beyond the training dataset. A group of methods named Physics-Informed Neural Networks (PINN) (Gasmi and Tchelepi, 2021; Chen et al., 2023; Gladchenko et al., 2023; Han et al., 2023) aim to incorporate the prior knowledge about the modeled system in the form of differential equations. These models are trained to simultaneously minimize the residuals of governing differential equations and the misfit with observational data. Recently, promising results were reported with approaches involving the use of Neural Operators (Badawi and Gildin, 2025) - a class of neural networks operating in infinite-dimensional function spaces.

Thus, most proxy modeling methods can be divided into two groups: CRM-like methods that use only integral parameters of the interwell environment. These models usually cannot properly characterize multilayer reservoirs and have difficulty describing nonlinear phenomena such as changes in fluid phase composition. Another group attempts to estimate reservoir dynamics with high accuracy, such as POD-based models and most ML methods (VAE-based models, PINN, Neural Operators). However, this accuracy comes at a cost. Some of these models such as those based on Galerkin or Petrov-Galerkin projections still require substantial computational resources for constructing and solving systems of equations, even in the reduced subspace. Others, as most ML-based methods, demand enormous amounts of data and resources to train complex model with numerous parameters, often struggling to generalize to scenarios that differ from those included in the training datasets.

In this work, a novel proxy modeling approach for reservoir simulation is proposed. In the proposed method, physical parameters such

as pore pressure and fluid phase composition are estimated in the cells perforated by wells. These values are used to estimate production rates using an approximation based on the Peaceman well model (Peaceman, 1978).

The proposed method aims for an intermediate position between the two groups described above overcoming some of their limitations. Estimation of the dynamics near the wellbore allows us to account for associated nonlinearities while keeping the number of model parameters relatively low, compared to models aiming to simulate high fidelity dynamics of the reservoir.

The paper proceeds as follows. Section 2 describes the proposed production rates estimation model. Section 3 contains results of the application of the proposed approach to estimate production rates for synthetic and real-world reservoir models as well as a comparison of the proposed approach with one of the variants of the CRM model. Section 4 discusses the obtained results, possible applications, and directions for further work.

2. Materials and methods

This work presents a complex approach to well flow modeling. The core idea behind the proposed method is to utilize the same technique for evaluating well flow rates that is used in conventional reservoir simulators: summarizing the inflow from the grid cells perforated by the well, whereas dynamic properties necessary for this evaluation -specifically fluid saturations and pore pressure - are estimated using machine learning-based models.

Fig. 1 shows the proposed workflow.

First, the bottomhole pressure (Pbh) and total injection rate (Qj°al = Qjn., where I is the set of all injector wells, Q^. - injection rate of the well j) are used to estimate the pore pressure values in the individual cells perforated by the well (p). The cell pressures and total injection rates are inputs to a separate model for estimating water and gas saturation (Sw, Sg) in the corresponding cell. Then, Sw and Sg are used to calculate the relative permeability (RP) values (kro, krw, krg). It is assumed that the dependencies of RP values on Sw and Sg are predefined in a manner consistent with reservoir simulation models. This can be achieved through parameterized correlations, such as Corey exponents, or by using tabulated values, where RP values are evaluated through interpolation of these tables. The same assumptions apply to PVT properties, such as formation volume factors, which are evaluated based on estimated pressure values in the corresponding cells. Flow rates (Q0, Qw, Qg) are estimated with a model based on the Peaceman well approximation using relative phase permeability, PVT properties, and pressure drops for individual cells. The pressure, saturation and flow rate models are discussed in detail in the continuation of this section.

2.1. Pressure model

The proposed approach for cell pressure estimation is based on the idea of pressure-pressure deconvolution (Onur et al., 2004, 2011). For a slightly compressible fluid with time-invariant fluid properties, the pressure distribution can be expressed as

p(r, t) = p0(r) + [ Pbh(t - x)hp(r, x) dx , (1)

J 0

where r is the spatial position vector, t is the time, p0(r) is the initial pressure and the position defined by r, PbhW = ^^ - first derivative of bottomhole pressure, hp(v,x) - response per unit change in bottomhole pressure (transient response) and x is the integration variable. Eq. (1) is valid in the case of only one source/sink, in the case of multiple sources/sinks the superposition principle should be applied. The idea behind deconvolution is to parameterize either the transient response or the impulse response (the response to a unit pulse of zero length, which is the derivative of a unit change in pressure) and evaluate it by

Fig. 1. Workflow of the proposed approach.

minimizing the deviation between the pressure estimate from Eq. (1) and the pressure measurements. However, if the fluid properties change (e.g., due to a change in phase composition), the linearity of Eq. (1) is not satisfied. The proposed approach involves the use of a convolutional neural network instead of explicit parameterization of the transient response. It mimics Eq. (1), but due to the introduced nonlinearities it does not reproduce it exactly and is able to handle nonlinear effects in the pressure response. Fig. 2 shows a schematic visualization of the model. The model maps the derivatives of the bottomhole pressure (Pbh) of the well and the total injection rate of all injection wells (Qtotal) to the pore pressure (p) dynamics in all cells perforated by the well.

In general, convolutional neural network contains one or several layers performing convolution operations. Input of such layer undergoes a scalar product with convolution kernel (filter) within a moving window. Each input channel is processed using a separate filter and the results are subsequently combined. Furthermore, each output channel corresponds to a distinct set of filters. Consequently, every convolu-tional layer is defined by nt x n0 filters, where nt is the number of input channels and n0 is the number of output channels. If no specific padding is applied to the input, the length of the output is given by lj — k + 1, where lj represents the length of the input and k denotes the size of the filter.

The convolutional models utilized in this work consist of three convolutional layers, each including a 1D convolution, with all layers except the last incorporating the Exponential Linear Unit (ELU) activation. The first input layer has two input channels (PbhW and Qtotal) and produces four output channels. After applying activation, these four channels are mapped to eight channels in the second layer, which are then mapped to the final output after another activation. The number of output channels corresponds to the number of grid cells perforated by the well.

One thing to be noted is that the standard convolutional window is symmetric around the position of the output value. In time series terms, this leads to the fact that the standard convolutional operation uses time steps from the past and future relative to the output value. In our work, we avoid this effect and shift the convolution so that we only use values from the past relative to the current position.

2.2. Saturation model

To estimate the relative phase permeabilities required to calculate flow rates, it is necessary to estimate the fluid saturation of all

f\total

PbPni

Fig. 2. Schematic representation of the pressure model. The blue vertical arrays contain time derivatives of the downhole pressure of the production well and the total injection rate of all injection wells at each time instant. Through a series of convolutions, the initial arrays are transformed into an output array (shown in green) containing the pore pressure in each perforated cell of a given well at each time instant.

perforated cells. Changes in fluid saturation are driven by changes in the level of water-oil and gas-oil contacts as a result of pressure changes and by fluid fluxes associated with the pressure gradient. e use a multilayer perceptron (MLP) model to estimate fluid saturation: a separate model is built for each perforated cell. The MLP is a neural network architecture composed of a collection of nodes (neurons) organized in a layered structure. Each neuron is connected to the outputs of all the neurons from the previous layer. It evaluates a linear combination of these outputs and applies nonlinear activation function to this combination:

=ni

■ wiAl + b'i

(2)

where a'. - value of the activation of the ith neuron of the Zth layer, W. . - weight of the connection between jth neuron of the (I — 1)-th layer and the ith neuron of the /th layer, b'i - bias of the ith neuron in the /th layer, f (■) is the nonlinear activation function. Weights and biases are adjusted during training procedure.

a

Table 1

Main parameters of numerical models.

Fig. 3. Saturation model scheme.

The scheme of the model is presented in Fig. 3. Inputs of the model are values of pore pressure in the cell ( ) and cumulative injection volume of all injectors of the oilfield ( Q? , = /0' Qtotal di). These values are processed through two layers consisting of four and three neurons with ELU activation and the last output layer consisting of two neurons with sigmoid activation f(x) = 1/(1 + e~x). The model maps input to water saturation (5"^) and gas saturation (Sg) values. The sigmoid activation in the output layer ensures that resulting saturation values remain bounded. Oil saturation ( S0) can be estimated as

^ — 1 - Su:

(3)

Saturation values are then used to evaluate relative phase permeabilities in order to estimate production rates.

2.3. Rate model

Sandface fluid inflow from the specific perforated grid cell to the well can be evaluated as (Abou-Kassem et al., 2006)

= (P

К

-P'f •

(4)

where subscript p e {o,w, g} denotes fluid phase: oil, water or gas correspondingly, - coefficient incorporating cell geometry, well trajectory and rock permeability (connection factor), - fluid viscosity, kr - relative phase permeability, p - pore pressure in the cell, pwf -flowing well pressure opposite to the cell, superscript denotes cell index. Assuming that the density of the fluid in the borehole is constant over time, we can express flowing well pressure opposite to the cell as

P'wf — Pbh + b'

(5)

where b' is a corresponding hydrostatic correction. The well flow rate at the surface is evaluated as the sum of the fluid inflows from all perforated cells, adjusted for surface conditions. Using Eqs. (4) and (5), and taking into account gas dissolved in oil and changes in fluid density, well production rates can be expressed as

K\

Qo — Zc' (p' - Pf

Qw — 2

(p ' - pf.

KB'o

KU

(p- (p- p'f ßg-g ßo-o У

(6a) (6b) (6c)

Main numerical model properties Values

Number of cells, XYZ 15 x 15x2

Reservoir dimensions, m 1500 x 1500 x 10

Initial pore pressure, bars 300

Vertical permeability, mD 1

Lateral permeability, mD 50

Simulation time, years 10

Number of time steps 120

Fig. 4. Pressure map on 90th time step.

where Bp - formation volume factor of corresponding phase, Rs - the ratio of gas volume dissolved in oil to oil volume in surface condition.

Cell pressure p' is evaluated using pressure model (Section 2.1), fluid properties (Rs, ц, B) are evaluated from PVT tables and estimated cell pressure. Fluid saturations are estimated with corresponding models and are used to evaluate relative phase permeabilities (kr). Thus, only estimates of the coefficients c' and V are required to estimate the flow rates using Eqs. (6). These values can be assessed by minimizing the residual between flow rate data from training scenarios and their estimates using Eqs. (6).

3. Results

3.1. Synthetic reservoir model

First, we test the proposed approach on a synthetic model of a simplistic reservoir. A set of synthetic scenarios was generated based on a single-phase model of a homogeneous reservoir. A single vertical production well is located in the center of the reservoir and penetrates the entire reservoir thickness. The well is controlled by the bottomhole pressure bh. The static properties of the model are shown in Table 1.

Fig. 4 shows pressure distribution at the 90th step in one of the generated scenario. A total of 50 scenarios were generated: each scenario corresponds to a particular well regime (Pbh) schedule for 120 time steps of one month duration. The bh schedule is constructed according to the following scheme: the value is randomly drawn from a uniform distribution i/ps.loo] with a probability of change of 10% at each time step during the simulation. An example realization of Pbh schedule is shown in Fig. 5.

Since the generated scenarios are single-phase, the saturation model is redundant. Thus, the flow prediction is performed in two steps using the pressure (Section 2.1) and flow rate (Section 2.3) models without saturation model described in Section 2.2. Of the 50 scenarios, 35 are used to train the pressure and flow rate models, and the remaining 15 scenarios are used for testing.

Bottomhole Pressure, bars

40 60 80 Time, months

Fig. 5. Sample control schedule.

The time-averaged relative error over the entire test set is less than 1% for the pressure model and 2% for the rate model. Fig. 6 is provided to demonstrate the performance of the pressure and flow rate models. These plots show the estimated and true pressure and flow rates for a scenario from the test set.

To study the performance of the pressure model in more detail, the pressure transient responses (h(t)) in one of the perforated cells obtained with the considered ML model and the industrial-grade reservoir simulator were compared. To obtain these responses, the _Pbh of the producing well was kept constant after the start of production. Thus, the transient pressure response can be evaluated as

X0 - Po

h(t):

(7)

Pbh - Po

where p(t) - is the pressure in the cell at time t, p0 is initial pore pressure in the cell, _Pbh is the constant bottomhole pressure.

Fig. 7 shows these pressure transient responses on a log-log-scale, as well as the response of the averaged pressure in the boundary cells of the reservoir model obtained with the reservoir simulator. It is seen that the pressure transient response obtained with the proposed pressure model agrees well with the response calculated with the reservoir simulator. The change in slope of the response after about 2 years of production is in good agreement with the onset of rapid pressure drop at the boundaries, demonstrating the sensitivity of the proposed models to such effects.

3.2. Real-world reservoir model

3.2.1. Training dataset

To train the models, a set of scenarios based on a real-world reservoir model was generated. The reservoir represents a complex layered structure saturated with oil, water and gas, developed using 16 production wells and 18 injection wells. The reservoir model consists of 6 million grid cells, 1 million of which are active. Fig. 8 represents the structure of the reservoir and the locations of the wells.

Training scenarios were generated by varying the well control schedule in the reference scenario. We utilized so-called restart models to create the training dataset. The state of the reference model as of January 2013 serves as the initial state for the generated training scenarios, which span from January 2013 to January 2027. Production wells are considered to be controlled by bottomhole pressure and injectors are considered to be controlled by injection rates. Series of well control values (bottomhole pressure) were generated by adding a multiplicative noise:

Kh«= <(0(1-

k(w - 0.5)) ,

(8)

from uniform distribution V[o,i]. Given the larger scale and structural complexity of the model compared to the simplistic synthetic model discussed in Section 3.1, we generated a total of 100 scenarios (as opposed to 50 for the synthetic model). For each scenario, k was randomly sampled from the set {0.1,0.2,...,0.9,1}. Fig. 9 shows the bottomhole pressure for one of the wells in one of the randomized scenarios and compared to the reference scenario.

Initially, the set of 100 scenarios was intended to have the same split between training and test datasets (70% and 30%, respectively) as for the synthetic reservoir model (Section 3.1). These scenarios were simulated using an industrial-grade reservoir simulator. The simulation was finished only for 78 scenarios, the rest simulations were aborted due to convergence problems. To maintain the size of the training dataset, the proportion of the test set was reduced. Consequently, the 78 completed scenarios were divided into 70 training scenarios and 8 test scenarios. Fig. 10 shows the production rates obtained after simulation of the scenario shown in Fig. 9.

3.2.2. Pressure model

To train the pressure models (Fig. 2), the results of training scenarios runs calculated on the reservoir simulator are used. The input data of the model are the time derivatives of the bottomhole pressures of the well and the total production rate of all injection wells, and the output data are the pore pressure values in the cells perforated by the well.

It has been observed that for some wells, the pressure dynamics in specific cells can be quite different, and a single model is often insufficient to correctly reproduce the pressure dynamics across all cells perforated by the well. Fig. 11 shows the pressure dynamics in different cells perforated by a well, where different colors represent pressure values in various cells. To improve the accuracy of the prediction, we divide all cells perforated by the well into groups so that the pressure dynamics in cells of the same group are similar, and train a separate model for each group. For this division, we sort all cells perforated by the well along the trajectory and evaluate the correlation coefficients between the pressure series in neighboring cells:

^ÛÂ^f(4+1- pMf

(9)

where ^bf - well bottomhole pressure in the reference scenario, k - coefficient responsible for variance, m is a random value drawn

where riJ+1 - correlation coefficient between pressure series in the ith and (' + 1)-th cells, p'. - pore pressure value in the 'th cell at the jth timestep, p' - mean pore pressure in the ith cell. Correlation coefficients are calculated for each training scenario and averaged over all scenarios. Fig. 12 shows the correlation coefficients between pressures in adjacent cells perforated by one of the wells.

We assume that the correlation coefficient below 0.9 characterizes the beginning of a new group of cells. Fig. 13 represents the pressure dynamics in different groups of cells perforated by a single well. Each subfigure illustrates the pressure dynamics for a specific group of cells. Each line represents the pressure dynamic of an individual cell, with colors corresponding to ones presented in Fig. 12.

To train the pressure models, 70 scenarios were randomly selected from the generated dataset, with only data prior to January 2023 used for training. The remaining 8 scenarios and the data after January 2023 were kept for testing. Each well was divided into groups of perforated cells as described above, and a separate model was built for each group. Training was performed by minimizing the MSE loss between the estimated values and the pore pressure values in the corresponding cells obtained from the full-scale reservoir simulator run. Fig. 14 represents the pressure estimates in cells perforated by the same well, but from different groups, compared to the values obtained from the full-scale simulator for one of the test scenarios.

¡+1

pi+1

Pj - P

P

r

i.i+1 -

Predicted pressure Actual pressure BHP schedule

40 60 80

Time, months

40 60

Time, months

(a) Pressure model performance.

(b) Production rate model performance.

Fig. 6. Proxy-model performance: (a) - pw( schedule (green line), predicted drawdown (blue line) and true drawdown (orange line) for a test scenario, (b) - predicted production rate (blue line) and true production rate (orange line) in the same scenario.

io>

Log of time, months

Fig. 7. Estimated unit transient response (orange line), unit transient response from numerical simulator (blue line) and mean boundary pressure (red line).

Fig. 8. Real-world reservoir model. Production wells are shown in green, while injection wells are shown in blue. The coloring of the grid blocks corresponds to porosity distribution.

3.2.3. Saturation model

To train a saturation model (Fig. 3), we use data from the same scenarios as for training pressure models. The same 70 scenarios are used for training and 8 scenarios are used for testing. Data after January 2023 are not included in the training dataset and were only used for testing. The inputs to the model are the pore pressure in the cell and the

140

120

100

k.

-Q 80

a. 60

40

20

0

Reference Scenario Randomized Scenario

2016 2018

2020 2022 Date

2024 2026

Fig. 9. Bottomhole pressure for one of the wells in one of the randomized scenarios (orange line) and reference scenario (blue line).

total injection volume for all injection wells in the field, and the outputs are the water and gas saturation in the corresponding cell. A separate model is trained for each perforated cell. We normalize the training data for each model so that the inputs (p and Qc. ) and the targets ( Sw and S) have zero mean and unit variance. Fig. 15 represents the estimated water and gas saturation dynamics in one of the perforated cells for the test scenario compared to the values obtained during the full-scale model run. The evaluation was performed using both the true pressure values in the cell and the values estimated by the pressure model (Fig. 2).

3.2.4. Rate model

To tune the flow rate models (Eqs. (6)) for specific wells, it is necessary to find such sets of values {.} and {bi} for each well that the flow rate values estimated by (6) are close to the values recorded during runs of the full-scale simulator. The cell pressure values ( ) obtained during full-scale simulator runs, as well as formation volume factor coefficients (B0, Bw, Bg) and Rs values calculated from PVT tables, are used for model training. To estimate the relative phase permeabilities (kro, krw, krg) saturation values estimated using models previously trained on true pressure values (see Section 3.2.3) are used. Adjustment of the model coefficients for each well is performed by solving the constrained optimization problem with the following loss function

*=11

i=1 j=1

Qo - QO +

-QU

2\

Qo + Qu

—2 Qi

(10)

where index ' represents a training scenario number, index j - a timestep number, Q0, Qw, Qg - oil, water, and gas production rates

- production - mean values

obtained during full-scale simulator run, Q0, rates estimation obtained with Eqs. (6), Q0

Qw and Q

Обратите внимание, представленные выше научные тексты размещены для ознакомления и получены посредством распознавания оригинальных текстов диссертаций (OCR). В связи с чем, в них могут содержаться ошибки, связанные с несовершенством алгоритмов распознавания. В PDF файлах диссертаций и авторефератов, которые мы доставляем, подобных ошибок нет.