Generalization of ML model (industriell forskning)

Offentlig prosjektsammendrag

A key component of petroleum production is insight into the state of individual wells. This is essential for safe and efficient operations, and essential to optimal production. The flowrate of gas, oil, and water through the well is one of the most important pieces of information about the wells. However, these flow rates are notoriously hard to measure, and the available solution can be prohibitively expensive, both in terms of equipment cost and the need for work intensive maintenance by a highly skilled workforce. With the emergence of better data infrastructure, more measurements, and the availability of data analysis tools, the community has opened up for data-driven alternatives to the traditional physics-based models. But, to the best of our knowledge, there has yet to emerge a solid data-driven virtual flow meter offering. Solution Seeker is proposing a new approach to data-driven virtual flow metering based on the concept of transfer learning. This enables us to utilize data from multiple wells from different assets to train a universal model that can be used for all wells with only minor adjustments. For instance, this means that wells with few data points can benefit from the universal knowledge from all known wells. This concept has seen great success within image and text analysis, but has received less attention in industrial applications. Solution Seeker believes transfer learning is key to developing a robust and sustainable data-driven virtual flow meter.