Banca de QUALIFICAÇÃO: JÉSSICA ALVES BRASIL

Uma banca de QUALIFICAÇÃO de MESTRADO foi cadastrada pelo programa.
STUDENT : JÉSSICA ALVES BRASIL
DATE: 11/04/2022
TIME: 08:30
LOCAL: meet.google.com/epj-ephy-gxp
TITLE:

DIAGNOSIS OF OPERATING CONDITIONS OF THE ELECTRICAL SUBMERSIBLE PUMP USING MACHINE LEARNING

KEY WORDS:

Electrical Submersible Pumps, Machine Learning, Classification and Operating Conditions.

PAGES: 83
BIG AREA: Engenharias
AREA: Engenharia Química
SUBÁREA: Tecnologia Química
SPECIALTY: Petróleo e Petroquímica
SUMMARY:

In artificial lifting, automation methods are used to increase the efficiency and production of oil wells. In an Electrical Submersible Pump (ESP) this becomes paramount, since the analysis of the available data is still insufficient to monitor, diagnose, interpret and analyze the performance and integrity of the well, in addition to the operation of the ESP and efficiency in real time. However, even though most of these wells operate with automated systems, due to the lack of early diagnosis of operating conditions, several problems can occur, including increased downtime, increased operating costs, decreased efficiency and, consequently, production losses. In practice, the activity of diagnosing operating conditions in ESP is usually performed by operators from the observation of patterns in amperimetric charts. However, difficulties such as the high number of wells per field and the lack of experience can reduce the effectiveness of this assignment. Currently, the ease of implementing classification techniques via Machine Learning and several published works on the subject, have challenged and encouraged companies to create solutions for early diagnosis of abnormalities in well operation. Thus, this work aims to provide a proposal for the detection of problems related to gas interference and gas blockage of the ESP pump from the analysis of electrical current data obtained for 20 wells in Mossoró, RN, Brazil. Machine Learning classification algorithms were implemented in the Python programming language in the Google Collaboratory environment. The classification algorithms used (with and without hyperparameter tuning) were decision tree, Support Vector Machine (SVM), K-Nearest Neighbor (KNN) and Multi-Layer Perceptron neural network. Balanced and unbalanced datasets were also tested. The results presented throughout the work confirm that the application of the ML algorithm is viable for the classification of the operating conditions presented, since all had an accuracy greater than 87%, with the best result being the application of the SVM model that reached an accuracy of 93%. Early-stage identification and resolution of ESP issues can lead to major cost savings and fewer maintenance requirements due to this smart system.

BANKING MEMBERS:
Presidente - 1149554 - OSVALDO CHIAVONE FILHO
Externa ao Programa - 6350734 - CARLA WILZA SOUZA DE PAULA MAITELLI
Externo ao Programa - 1752014 - EDNEY RAFAEL VIANA PINHEIRO GALVAO