Identificação de modelos Narx para poços surgentes de petróleo usando programação genética multi-gênica

Abstract

Artificial intelligence techniques applied in real contexts such as prediction of process variables, detection of anomalies or changes in production lines can result in reduced maintenance costs, actions to prevent accidents and failures, decision-making support, avoid losses of production and financial aspects and identify points for process improvement. The text presents a proposal for the application of a soft sensor, using an evolutionary algorithm, to identify future behavior of a temperature sensor during the operation of oil wells. The approach involves multiple executions of the multi-gene genetic programming algorithm (MGGP) to identify hyperparameter values contributing to the best results. After analyzing and determining the parameters for MGGP, the soft sensor, based on a polynomial NARX model, is implemented on a real public dataset called 3W. This dataset contains sensor measurements from oil wells between the periods of 2017 and 2018. The performance of the soft sensor is evaluated using the Mean Absolute Percentage Error (MAPE), which indicates the model's adequacy through the relative error between real data and model output expressed as a percentage. The model is trained and validated on data from two available wells, showing satisfactory results in representing the normal operation dynamics of each oil well. However, during periods of measurement transitioning to an anomalous state, the model is unable to explain such behavior. As it evolves through the transitional period, the error tends to increase. This outcome suggests the model's potential use in the context of anomaly detection in oil wells, functioning as a one-class classifier.