Enhanced Neural Network Architectures for Data-Scarce Environments and Multi-Parameter Prediction in Oil and Gas Operations

Abstract

Neural networks are a crucial component of modern artificial intelligence, demonstrating impressive abilities in understanding complex patterns and relationships in data. However, applying neural networks in real industrial systems, such as Oil and Gas operations, is challenging due to limited his- torical data availability, especially for new machines, and the high cost of obtaining or producing data. As a result, there is a scarcity of public data for the research community. This PhD thesis proposes innovative neural net- work architectures tailored to address the critical challenges in this field. To solve the issue of low prediction accuracy in predicting the health state of tools, a novel neural network architecture is proposed to forecast the Remain- ing Useful Life when limited training data is provided. a feedback mechanism is incorporated into an artificial neural network in a novel manner, using the values of the output layer neurons as inputs. These inputs are utilized as features to generate precise predictions. To validate the effectiveness of the approach, real dataset from oil and gas wells during production is used, this study focuses on a sub dataset of a sub-surface safety valve tool. Addition- ally, a custom neural network architecture is proposed to create a data-driven digital twin based on multi-target regression to mitigate the time delay that impacts decision-making for drillers during directional drilling operations. The architecture combines Long-short Term Memory and Multi-Layer Per- ception branches in a single neural network to forecast and predict important drilling parameters, such as inclination and rate of penetration. To validate this approach a real data collected during a directional drilling operation is used. Furthermore, an incremental learning framework is implemented to simulate the performance of the architectures in real-time, where data is continuously received and the regression models are updated concurrently. The proposed architectures demonstrate superior results compared to exist- ing works in the field. The research conducted in this thesis aims to extend the capabilities of neural network models, uncovering their potential in solv- ing complex problems while contributing to the evolving field of intelligent systems.