DEEP LEARNING FOR ESTIMATION SYSTEM OF MULTIPHASE FLOWRATES OF OIL AND GAS PRODUCTION WELLS

• Main Author: Nurmalia, Enung
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/81483
• Institution: Institut Teknologi Bandung

Artificial intelligence, evolving within the models of deep learning and machine learning, has shown widespread applications in various fields. One of the industries that has increasingly embraced the potential of artificial intelligence is the oil and gas sector. The oil and gas industry confronts numerous complex challenges, one of the most significant being the analysis of production wells with multiphase flow. Multiphase flow encompasses the simultaneous movement of different phases, such as oil, gas, and water, within a single stream from a production well. This intricate mixture requires advanced technology and highly reliable measuring equipment to accurately capture and interpret the data. Managing these complexities demands substantial investments in the procurement, installation, and maintenance of sophisticated instruments. To address these challenges, the use of multiphase flowmeters (MPFM) has become widely adopted. Production wells with a gas volume fraction (GVF) exceeding 95%, it is more appropriate to use a wet gas meter (WGM) type of MPFM. WGMs have low measurement uncertainties at ±5% for oil and gas hydrocarbons. The high accuracy provided by WGM technology is critical for enhancing daily operational efficiency and optimizing production and for advancing research and development. The oil and gas industry depends on a wide array of measurements across all phases of its operations, from the wellhead and flowlines to processing facilities and sales areas. These measurements provide critical real-time data for monitoring, control, and safety systems, and are also stored as part of the industry's historical data reserves in historian servers. This extensive dataset can be mined to extract valuable insights, particularly through the application of artificial intelligence. It is increasingly utilized to predict multiphase flow rates due to its prowess in handling complex, nonlinear systems and its capability to analyze large datasets efficiently. By offering accurate and reliable predictions and diminishing the need for physical measuring devices, artificial intelligence helps lower operational costs. These advantages position artificial intelligence as an effective and precise tool for enhancing operational performance and supporting more informed decision-making. The integration of deep learning and machine learning into this research is to explore the ability of these artificial intelligence models to handle complex and non-linear systems, leverage large dataset, provide accurate predictions, and reduce costs as it reduces the needs of physical models. These artificial intelligence algorithms will be employed to learn and estimate production well rates using data from the WGM and from other instruments deployed around the production wells. This approach holds the potential to enhance the accuracy of well production rate estimations, making it applicable in real-world scenarios such as production estimation, and well testing. In this research, deep learning models employ Long Short-Term Memory (LSTM) and Deep Neural Network (DNN). Meanwhile the machine learning models employ random forest, extreme gradient boosting (XGBoost), light gradient boosting machine (lightGBM), support vector regression (SVR), and k-nearest neighbors (kNN). The modeling performance is assessed through key metrics, including mean absolute error (MAE), root mean squared error (RMSE), mean absolute percentage error (MAPE) and coefficient of determination (R_squared). The validation of the machine learning models is evaluated by cross validation with k-fold method. Uncertainty quantification is applied to deep learning models to assess the uncertainty and reliability of models using Monte Carlo Dropout and Ensemble methods. The parameters employed in this research originate from the real operation of oil and gas production wells, encompassing crucial variables for modeling such as wellhead pressure, wellhead temperature, topside pressure, choke valve opening, and parameters gauged by the WGM, specifically pressure, differential pressure, and temperature of fluid at flowline, and computational results include gas and oil flowrates. The employed dataset spans a 5-year timeframe, ranging from May 2018 to October 2023 with 10 minutes sampling time, involving 262,630 data points. It is observed from the simulation results that for deep learning algorithms, DNN dominated the performances of model fitness, minimum errors and low uncertainty for estimating the two targets of gas and oil flowrates. It means that the models performed with high accuracy and reliability with the value of R2 98.91%, MAE 0.29%, MAPE 1.95% and uncertainty 2.17%. The variations of data segmentation and hyperparameters were applied to these deep learning models to get better comparisons and insight about how well the models adapted to different inputs and parameters. Meanwhile for machine learning models, random forest exhibited superior performance for estimating the gas and oil flowrates across different data segment variations with the value of R2 99.48%, MAE 0.4%, MAPE 1.0%, mean RMSE from cross validation at 0.8% with standard deviation of RMSE at 0.55%. Keywords: deep learning, machine learning, multiphase flowrate, multivariate time series, oil and gas production well, uncertainty quantification, wet gas meter ?