Abstract The downhole pressure is an important variable used to optimize the oil production in deep-water oil wells. However, due to its localization at the seabed, its sensor breaks down easily. Thus, a parameter-less Evolutionary Algorithm, called Evolutionary Algorithm with Numerical Differentiation (EAND), is proposed in this work for designing soft sensors to predict the downhole pressure. Results show that the EAND performs good balance between local and global searches, providing the best results in 17 out of the 20 optimization problems, and achieving the fastest convergence in 16 simulated problems. The proposed algorithm yielded the best soft sensors under the five offshore oil wells studied when compared to other identification methods. Three kinds of nonlinear models for prediction were implemented, and ensembles composed of decision trees (Random Forest) obtained the best results. The Mean Absolute Percentage Errors (MAPE) found when predicting the downhole pressure by the identified soft sensors ranged from 0.1453% to 0.788%, which are very satisfactory.
Paper
Full text
Soft sensors design in a petrochemical process using an Evolutionary Algorithm
Semantic Scholar · Engineering · 2019
Abstract
Abstract The downhole pressure is an important variable used to optimize the oil production in deep-water oil wells. However, due to its localization at the seabed, its sensor breaks down easily. Thus, a parameter-less Evolutionary Algorithm, called Evolutionary Algorithm with Numerical Differentiation (EAND), is proposed in this work for designing soft sensors to predict the downhole pressure. Results show that the EAND performs good balance between local and global searches, providing the best results in 17 out of the 20 optimization problems, and achieving the fastest convergence in 16 simulated problems. The proposed algorithm yielded the best soft sensors under the five offshore oil wells studied when compared to other identification methods. Three kinds of nonlinear models for prediction were implemented, and ensembles composed of decision trees (Random Forest) obtained the best results. The Mean Absolute Percentage Errors (MAPE) found when predicting the downhole pressure by the identified soft sensors ranged from 0.1453% to 0.788%, which are very satisfactory.