Abstract
The phase and volumetric behaviour of reservoir fluid properties, referred to as pressure-volumetemperature (PVT) properties, involve the thermodynamic studies of the fluid with respect to pressure, temperature and its volumetric compositions. PVT properties are usually determined by laboratory experiments performed on the actual samples of the reservoir fluid. Failing that, these fluid properties have been evaluated by some other methods such as equations of state, empirical correlations and recently, machine learning models.Machine learning is basically the prediction of the future with, (supervised learning), or without, (unsupervised learning), prior knowledge of the past. A common problem for the standalone machine learning technique is local minimum. In view of this, ensemble systems and hybrid techniques have been developed successfully for improvement in different fields.
This work introduces two different ensemble methods based on support vector regression and regression trees where both ensemble approaches utilise a novel concept tagged “Tying Ranking” in selection of the base models. Also, a hybrid system for reservoir fluid characterisation with a novel way of grouping petroleum fluid properties using intelligent method was developed. The hybrid system uses K-Means clustering for the intuitive grouping along with functional networks for the prediction.
The performance and generalisation of the developed models are compared against their standalone and selected empirical models using some statistical measures which are commonly used for performance evaluation in the petroleum industry.
In the first category of experimentation, the impact and effect of training the machine learning models with more diverse and bigger data set is shown. Effects of using different functional forms to predict dead oil, saturated and undersaturated viscosity are also explored. In addition, impacts of different statistical measures on the predicted outputs and wrong interpretations of results in the literature are examined.
The main statistical measures that are used for comparison are root mean squared errors, average absolute percentage relative error and maximum absolute percentage relative error. For each of the reservoir fluid properties considered in this work, at least one or more of the developed machine learning models have better overall and average performance than all the compared correlations in each category.
The superiority of the three developed machine learning models is visible in the trend analysis as they show less deviations in results compared to the empirical correlations and their standalone methods in most cases for all the considered reservoir fluid properties.
Date of Award | Jun 2018 |
---|---|
Original language | English |
Awarding Institution |
|
Supervisor | Mohamed Hassan Sayed (Supervisor), Mohamed Bader-El-Den (Supervisor) & James Buick (Supervisor) |