NeuralCompass
The well flow rate estimation problem is at the heart of true production optimization. Unless we know what each well is producing at any given time, optimization in real-time becomes impossible. However, in a recent study we did in-house, we found that existing VFM and MPFM solutions in the market typically achieve an error of 20% or less for only 50% of all test points. Naturally, the well-maintained solutions achieve a greater performance but with typically high maintenance costs.
Over more than 10 years, our team has worked on developing our data-driven VFM NeuralCompass (NC). The solution utilizes state-of-the-art machine learning combining neural networks with multi-task learning, Bayesian inference, and online learning to offer the market high-quality flow rate estimates with 90% less effort. Using our platform ProductionCompass AI we take advantage of our Squashy data mining algorithms to ensure automatic calibration with high quality and relevant data.
- Live
-
Live updated rate estimates written back to your data historian as timeseries / tags, to be used in all your relevant workflows
- Scalable
-
Extreme scalability and maintainability
- Quality control
-
Specialized dashboards and tools in ProductionCompass AI for quality control and performance monitoring of the models
- Enables optimization
-
Enables real-time production optimization, with a large potential impact on recovery
Below you see an example of a high-fidelity NeuralCompass flow rate estimate together with auto-detected steady-state intervals in gray detected by our Squashy algorithm.
Our NC flow rate estimates can be used in well monitoring, providing situational awareness for engineers in between well tests. NC-VFM can be used as a stand-alone product or as a backup system to an MPFM. The flow rates are also appropriate to use in downstream applications such as allocation, production optimization, or reservoir modeling. In the following section, we will deep-dive into our allocation service FlowFusion. Although we recommend using NC-VFM in FlowFusion, the service is fully operable without NC-VFM using rate estimates that you already have access to such as MPFMs or other VFMs.
Our research and development in this space is happening at great speed, and all our existing and potential clients can expect great advancements in this area. We have published several papers on the topic, with yet more in the pipeline. You can find our newest publications here:
- Multi-task learning for virtual flow metering
- Link
- Bayesian Neural Networks for Virtual Flow Metering: An Empirical Study
- Link
- Multi-unit soft sensing permits few-shot learning
- Link
- Passive learning to address nonstationarity in virtual flow metering applications
- Link
- Sequential Monte Carlo applied to virtual flow meter calibration
- Link
- Multi-task neural networks by learned contextual inputs
- Link
Get in touch to hear more and see our latest results!