Reservoir Characterization and Prediction Modeling Using Statistical Techniques (2017-2021)
Investigator(s): Halldora Gudmundsdottir
Reservoir characterization and prediction modeling have long been among the more challenging tasks in geothermal reservoir engineering. The main reason is the presence of fractures and faults, which control the mass and heat transport in the subsurface. Nowadays, with a substantial increase in data due to advances in computer power and measuring equipment, the oil and gas as well as the geothermal industries are presented with some of today’s most complex data science problems. Therefore, statistical methods are becoming increasingly popular as tools for predictive analysis in the exploration, production and delivery phases.
In this work, the applicability of using statistical methods for reservoir characterization as well as prediction modeling was explored. Three methods were analyzed and applied on a synthetic library of fracture networks. First, the Alternate Conditional Expectation (ACE) algorithm was used to estimate well-to-well connectivity between injection and production wells using tracer return and temperature data. Second, k-means clustering was applied where fractures of similar character were grouped together and interwell connectivity and thermal behavior estimated. Third, a Canonical Functional Component Analysis (CFCA) was used to directly forecast thermal behavior along with uncertainty quantification.
For the ACE algorithm, the results obtained with tracer data were in good agreement with tracer transit times, for 80.5% of the fracture networks the ACE connectivity was within ±0.05 of the connectivity implied by transit time, while temperature data showed much less correlation to connectivity with the ratio reduced to 58.3%. K-means clustering showed promise in being able to group fracture networks in agreement with connectivity between wells. These groups could then be used to predict the temperature of wells using new data. However, the complexity grows when more producers are added, and therefore the number of groups increases in order to adequately describe the character of the fracture networks. The CFCA method involves building a statistical relationship between historical and forecasting data and using regression to quantify forecast uncertainty. Applying CFCA on our data, we were able to make direct forecasts for the producers based on tracer or temperature data as well as significantly reduce predicted ranges of thermal responses for the wells. The CFCA method assumes data is smooth and its success depends on finding a linear relationship between historical and forecast data in the canonical space. Therefore, introducing real data, this relationship will most likely become less linear due to added noise, and meaningful predictions harder to obtain.
References:
Gudmundsdottir, H and Horne, R. “Reservoir Characterization and Prediction Modeling Using Statistical Techniques.” Proceedings: 43rd Workshop on Geothermal
Reservoir Engineering, Stanford University, Stanford, CA (2018).