Development of six machine learning models for pump intake pressure calculations in esp wells /
Mohamed Hamdy Mohamed Ibrahim El-Sersy,
Development of six machine learning models for pump intake pressure calculations in esp wells / تطوير ست نماذج للتعلم الآلي للتنبؤ بضغط سحب المضخة في الابار التي تستخدم طرق الرفع بالمضخات الغاطسة / By Mohamed Hamdy Mohamed Ibrahim El-Sersy; Under the Supervision of Prof. Dr. Mohamed Helmy Sayyouh, Prof. Dr. Ahmed Hamdy El-Banbi, Prof. Dr. Mahmoud Abu El Ela - 117 pages : illustrations ; 30 cm. + CD.
Thesis (M.Sc.)-Cairo University, 2023.
Bibliography: pages 66-69.
The pump intake pressure (PIP) is a crucial parameter for optimizing
the performance improvement of pumped oil wells. Recently, Electrical
submersible pump (ESP) systems have usually adopted downhole gauges to
evaluate PIP. In this study, six new models (Linear Regression (LR) model,
Polynomial Regression (PR) model, Decision Tree (DT) model, Random Forest
(RF) model, Support Vector Machine (SVM) model, and finally, Artificial Neural
Network (ANN) model) were developed based on machine learning algorithms
and Artificial Intelligence (AI) techniques to predict pump intake pressure (PIP)
from readily available data, and field measurements in ESP pumped wells.
A database of 2352 field data points was collected from 105 ESP wells to
develop the six new models. The data was split into 60% (1411 data points) for
training and 40% (941 data points) for testing. The developed models rely on
the following measurements as input parameters: wellhead pressure, total
production rate, water cut, oil gravity, pump setting depth, the net liquid above
the pump, tubing size, casing size, and casing pressure.
The models' accuracies were compared against each other. Then, the
developed models were compared against the accuracy of previous
correlations and actual readings obtained from ESP downhole pressure
sensors. The results indicate that the accuracy of the ANN model is significantly
higher than that of the other developed machine learning models and the
previously available correlations. Using the ANN model, the Average Mean
Absolute Error (AMAE) comparing the calculated and the measured PIP is
11.93% and 12.33% for the training and testing data, respectively. These
results demonstrate the strength of the developed model to predict the PIP with
better accuracy and without the need for a downhole pressure sensor. تم استنباط ستة نماذج جديدة وهي نموذج الانحدار الخطي ونموذج الانحدار متعدد الحدود ونموذج شجرة القرار ونموذج الغابة العشوائية ونموذج آلة المتجهات الداعم وأخيراً نموذج الشبكة العصبية الاصطناعية. تم تطوير هذه النماذج استنادًا إلى خوارزميات التعلم الآلي وتقنيات الذكاء الاصطناعي، للتنبؤ بضغط سحب المضخة من القياسات الحقلية المتاحة بسهولة وبيانات الآبار التي تستعمل المضخات الغاطسة الكهربائية
Text in English and abstract in Arabic & English.
Petroleum Engineering
Artificial Intelligence ESP wells Artificial Neural Networks
622.18
Development of six machine learning models for pump intake pressure calculations in esp wells / تطوير ست نماذج للتعلم الآلي للتنبؤ بضغط سحب المضخة في الابار التي تستخدم طرق الرفع بالمضخات الغاطسة / By Mohamed Hamdy Mohamed Ibrahim El-Sersy; Under the Supervision of Prof. Dr. Mohamed Helmy Sayyouh, Prof. Dr. Ahmed Hamdy El-Banbi, Prof. Dr. Mahmoud Abu El Ela - 117 pages : illustrations ; 30 cm. + CD.
Thesis (M.Sc.)-Cairo University, 2023.
Bibliography: pages 66-69.
The pump intake pressure (PIP) is a crucial parameter for optimizing
the performance improvement of pumped oil wells. Recently, Electrical
submersible pump (ESP) systems have usually adopted downhole gauges to
evaluate PIP. In this study, six new models (Linear Regression (LR) model,
Polynomial Regression (PR) model, Decision Tree (DT) model, Random Forest
(RF) model, Support Vector Machine (SVM) model, and finally, Artificial Neural
Network (ANN) model) were developed based on machine learning algorithms
and Artificial Intelligence (AI) techniques to predict pump intake pressure (PIP)
from readily available data, and field measurements in ESP pumped wells.
A database of 2352 field data points was collected from 105 ESP wells to
develop the six new models. The data was split into 60% (1411 data points) for
training and 40% (941 data points) for testing. The developed models rely on
the following measurements as input parameters: wellhead pressure, total
production rate, water cut, oil gravity, pump setting depth, the net liquid above
the pump, tubing size, casing size, and casing pressure.
The models' accuracies were compared against each other. Then, the
developed models were compared against the accuracy of previous
correlations and actual readings obtained from ESP downhole pressure
sensors. The results indicate that the accuracy of the ANN model is significantly
higher than that of the other developed machine learning models and the
previously available correlations. Using the ANN model, the Average Mean
Absolute Error (AMAE) comparing the calculated and the measured PIP is
11.93% and 12.33% for the training and testing data, respectively. These
results demonstrate the strength of the developed model to predict the PIP with
better accuracy and without the need for a downhole pressure sensor. تم استنباط ستة نماذج جديدة وهي نموذج الانحدار الخطي ونموذج الانحدار متعدد الحدود ونموذج شجرة القرار ونموذج الغابة العشوائية ونموذج آلة المتجهات الداعم وأخيراً نموذج الشبكة العصبية الاصطناعية. تم تطوير هذه النماذج استنادًا إلى خوارزميات التعلم الآلي وتقنيات الذكاء الاصطناعي، للتنبؤ بضغط سحب المضخة من القياسات الحقلية المتاحة بسهولة وبيانات الآبار التي تستعمل المضخات الغاطسة الكهربائية
Text in English and abstract in Arabic & English.
Petroleum Engineering
Artificial Intelligence ESP wells Artificial Neural Networks
622.18