Eventos Anais de eventos
COBEM 2023
27th International Congress of Mechanical Engineering
Data-driven multiphase flow parameters prediction capabilities and limitations on a real oil well production data
Submission Author:
Anderson Faller , SP
Co-Authors:
Anderson Faller, Saon Vieira, Bernardo Foresti, Adriano Todorovic Fabro, Marcelo Souza de Castro
Presenter: Anderson Faller
doi://10.26678/ABCM.COBEM2023.COB2023-1242
Abstract
Virtual Multiphase Flowmeters are model-based tools to estimate multiphase flow rates in pipelines that may replace physical flowmeters or test separators whenever those are not available, non existent or its use is not possible. They can be based on physics principles (simulation) or on data-driven models. We explore the application of the latter type across a few machine learning architectures by proposing a performance evaluation depending on the prediction goals. We state that data-driven models degrade over time due to changes in the operating conditions (short-term degradation) or due to slow changes in the fluid characteristics or reservoir inflow conditions (long-term degradation). Then, we apply the proposed methodology to quantify these degradations on a dataset extracted from a real oil well for over 3 years of its productive life. We show that, after training a model with a few hours of data, the prediction error increases on average 1-2 percentage points in the first 8 hours within the extrapolation range. When training a model with 2 years of historical data, the prediction error increases consistently with a rate of 30 percentage points in 10 months. We also show that larger models with time shifted input features yield better predictors.
Keywords
Petroleum, virtual flowmeter, machine learning, Virtual Sensing, oil production
