Predictive modeling with applications in decision support systems for oil and gas production

Abstract

In the past decade, artifical neural networks (ANN) has given us self-driving cars, practical speech recognition and more effective web search. It is in interest of petroleum industry to research the applicability of ANNs in decision support systems for the operation of oil and gas production systems. This interest coincides with increased availability of computational power and sensor data in the petroleum industry. This thesis investigates on using different feed forward ANN architectures and their applicability to virtual flow metering and steady-state production optimization applications. The predictive capability of the proposed models are assessed on steady-state production data acquired from an offshore production system located on the Norwegian continental shelf. The results presented in this thesis shows that feed forward ANNs can be used as virtual flow meters with accuracy comparable to modern virtual flow meters with average errors below 10%. Compared to physics based virtual flow meter models, feed forward ANNs are fast to build and maintain, and requires lower level of engineering skill, making it a cheaper alternative to physics based virtual flow models. The results shows that the proposed neural network model for production optimization do not have accuracy good enough to be used for production optimization. Production optimization involves exploring and evaluating many possible operating points and requires therefore better predictive capabilities than virtual flow models which the model could not provide. However, there is room for improvements and the results are therefore promising.