Automated Well Monitoring: Machine Learning and Web Application

Abstract

The challenge within the oil and gas industry is that of complexity and therefore cost, specifically due to the tough working environments and delays/downtime . Therefore, digitization is proposed as a cost saving opportunity ,making data collection through sensors and data analytic approaches priority in the industry. The integration of machine learning has already shown its contribution in augmenting human decision making and optimizing processes. However deciding the right technique keeps being a challenge for further reducing the cost. In this thesis, we are assembling several machine learning and deep learning models and testing them with the aim of optimally reconstructing missing flow rates in well monitoring data. The experiments are focused in some simple regression models such as Linear Regression, Ridge Regression, Kernel Ridge Regression and Gradient Boosting Regression and one deep learning model for time-series: LSTM. Except of its original form, we are applying two feature engineering techniques on the well data: convolution method and transient reduction method. The experimental results shows the importance of feature transformation in performance of the models by emphasizing two moments: the dramatic improvement of Kernel Ridge Regression over the convolutional transformed data and the outstanding ability of LSTM to learn on raw data. We finalize our work by creating a simple web application for reconstructing the missing flow rate values using the most optimal machine learning model, by giving a chance for everyone to reuse and interact with the model.