Tese e Dissertação

Tese: Virtual Flow Metering and Synthetic Well Logs Generation Using Multimodel Machine Learning Strategies and System Identification

Multiphasic flow metering and well logging are key information that helps oil and gas companies understand reservoir characteristics and make strategic decisions. Recently, machine learning models have become an alternative to traditional data acquisition procedures based on physical hardware or first principle models, since the latter are, in general, expensive, time-consuming, and often unreliable. In this work, we propose new approaches for machine learning models to enhance the performance of virtual flow metering models and synthetic well log generation. For that purpose, this dissertation develops three application cases based on actual well data from the Volve Field in the Nothern Sea. Firstly, the application of stacking ensembles and non-linear system identification techniques improved RMSE metrics by 4.5% to 29% compared to benchmark regressors. System identification proved to be an effective approach for assessing optimal lags while stacking ensembles enhanced well-rate prediction by integrating various base learners. Secondly, a two-step training approach, applying a residual predictor, improved the quality of the well rates predictions by 1% to 13% depending on the complexity of the base estimator. In the third case, seasonal decomposition was used whether as a feature extraction method or as an ensemble strategy for well log generation purposes. While the first promoted a 27% gain in RMSE for compressional slowness log, the second outperformed the baseline models for 6 of the 8 examined estimators, reducing the density log RMSE by 5.2%. In all four scenarios examined, both methods achieved better results than the baseline strategy. These findings provide new guidelines for research in the field of virtual flow metering and well logs generation as well as in other domains.

Link da defesa:

https://puc-rio.zoom.us/j/98410105940?pwd=SC5F2Vs6QY7ddxtcapXnOyJ28n8Reo.1

Ver todos