Disputas: Zhe Ban

Zhe Ban skal forsvare avhandlingen sin for graden philosophiae doctor (ph.d.) ved Universitetet i Sørøst-Norge. 

Hun har fulgt doktorgradsprogrammet i prosess, energi og automatiseringsteknikk ved Fakultet for teknologi, naturvitenskap og maritime fag. 

Alle interesserte er velkomne til å følge prøveforelsningen og disputasen.

• Les avhandlingen her

Sammendrag

This PhD work set out to address the challenging problem of data reconciliation and model fitting in dynamic nonlinear oil production systems. Both Monte Carlo-based numerical approaches and machine learning-based methods were investigated and applied in various scenarios.

Two types of oil production systems, including gas lifting oil wells and electrical submersible pump-lifted oil field, were considered for different scenarios. Measurement selection method was adopted to benefit efficient computation during estimation. With considering uncertainty, various Monte Carlo approaches were applied and compared for data reconciliation with simultaneous parameter estimation in both online and offline cases.

The proposed MCMC based method can track abrupt changes of multiple parameters as well as the uncertainty of these parameters in real-time without requiring empirical prior knowledge. This method was compared with various nonlinear filters for time-varying parameter estimation in oil production. The advantage of the MCMC scheme is that it is possible to add prior knowledge of the dynamic parameter during estimation at all time steps by including condition statements for samples. However, the UKF-based method has a shorter delay time, making it more efficient and accurate for real-time parameter estimation, especially in scenarios with continuously changing parameters.

Data reconciliation is primarily driven by first principles model-based methods. However, when there is a significant model mismatch or parts of the system are challenging to describe mathematically, traditional data reconciliation methods struggle to provide accurate estimations. Part of the PhD research concentrated on integrating first principles models with neural networks to produce accurate data- driven output estimations. Based on the results obtained, we can see that even when a part of a first principles model or a simplified physical model is involved, it can help machine learning methods improve their explainability, precision, and estimation speed. The proposed PINN model is able to predict more details of the dynamics, especially for transients.

All in all, throughout all the conducted work, this research demonstrates that various approaches can be chosen for the data reconciliation and model fitting of oil production systems in different circumstances. The proposed methods fill a gap in the research on dynamic modeling in oil production under uncertainty, which potentially benefit model-based control and subsequent real-time optimization.