Determination of oil well properties using fiber optic sensor (DAS) data and machine learning models

Abstract

In this Master dissertation data driven modeling of multiphase fluid fraction prediction is implemented. Fluid fraction is predicted for two phase flow, one for GVF and oil flow and second for Water cut and oil flowrate. The DAS fiber optic sensor data is collected from NORCE Lab.

The raw DAS data is processed by applying short-time Fourier transform and then processed images dataset is created. The feature extraction from images is performed using two different models, one is ResNet50 and other is custom CNN model. Then using different models such as random forest regressor, gradient boosting regressor, neural network regressor and support vector regressor model is trained for prediction of fluid fraction.

The predictions results are visualized in the form of scatter plots, correlation matrices, residual and in numeric form. Also, different models are compared and different types of datasets are used for training models.

Instead of a small dataset the prediction of the models is quite good with an R2 score of 0.70. Random forest regressor and gradient boosting regressor showed quite good results.