Machine learning assisted history matching to integrate fiber optic data with reservoir simulation
Document Type
Conference Proceeding
Publication Date
1-1-2020
Abstract
A novel automated workflow is presented for integrating fiber optic Distributed Temperature Sensing (DTS) data in Cyclic Steam Stimulation (CSS) reservoir simulations using a machine learning assisted history matching workflow. This study uses actual field data from a horizontal well CSS operation in a heavy oil field in California. The value of integrating DTS in the history matching process is illustrated as it allows the injection profile to be accurately estimated along the entire length of the well. Since the steam-oil relationship is the main driver for forecasting and decision making in thermal recovery operations, knowledge of downhole steam distribution across the well over time can decrease uncertainty in predictive modeling and optimize injection and production. High resolution large DTS dataset is integrated in numerical simulations for estimating downhole steam injection profile, by utilizing machine learning techniques. A stepwise grid-refinement approach was implemented to optimize computational efficiency and improve predictive accuracy. A multi-segmented wellbore model enabled the distinction between the wellbore dynamics and the reservoir effects on the flowrate and temperature response. The proxy models created from statistical and machine learning techniques streamlined the workflow and also helped to quantify the uncertainty due to the reservoir heterogeneity. Particularly, the Design Exploration Controlled Evolution (DECE) optimization engine enabled the entire range of uncertainty to be captured by applying machine learning to history match the water, oil, and temperature profiles simultaneously. DECE algorithm's ability to handle a large number of parameters was particularly useful in incorporating grid refinement to account for small-scale heterogeneities such as localized hard streaks and interbedded shale. In effect, a more complete history match is achieved by incorporating both production and DTS data, and steam conformance is more rigorously assessed. Visualization of the DTS thermal profiles through time also helped in identifying downhole issues in real-time. Particularly, thermal communication with another well was detected due to an unexpected increase in downhole temperatures in the test well. This also resulted in non-uniform reservoir heating and poor steam conformance. After the communicating well was worked over, an improved steam profile was observed from the DTS data. The continuous downhole monitoring using fiber optic surveillance made it possible to not just detect the thermal communication event in real-time, but also assess the effectiveness of the remedial workover, which could have been missed with conventional logging.
Publication Source (Journal or Book title)
Society of Petroleum Engineers - SPE Canada Heavy Oil Conference 2020, CHOC 2020
Recommended Citation
Feo, G., Sharma, J., & Cunningham, S. (2020). Machine learning assisted history matching to integrate fiber optic data with reservoir simulation. Society of Petroleum Engineers - SPE Canada Heavy Oil Conference 2020, CHOC 2020 Retrieved from https://repository.lsu.edu/petroleum_engineering_pubs/602
