Temperature monitoring is the most important surveillance in thermal assets, but temperature logging is limited in frequencies and locations. In addition, it is extremely difficult to review all the measured temperature and injection data manually since there are 10,000+ wells in Kern River field. To overcome the limitations, data-driven reservoir temperature models are presented that are built using past temperature logs and steam injection rates of the Kern River field, California.

Based on the physics and geologic understanding the reservoir, adequate input features were selected and queried. Data cleanup was conducted to remove erroneous data or fix data errors using statistical tools such as multivariate Gaussian distribution. Voronoi diagram based dynamic injector selection algorithm (DISA) was developed to correctly capture the injectors which impact on temperature changes of a temperature observation well. Based on geologic characteristics of the Kern River, reservoir was divided into two sub-reservoirs, North-East and South-West. Two full field models were developed for predicting maximum and mean temperatures of a heated zone with multi-layer perceptron for both sub-reservoirs using about 120,000 data points from over 25,000 temperature curves measured at 700+ temperature observation wells.

To estimate proper model update frequencies and verify the process, three yearly models (models 2015, 2016, and 2017) were built and validated by using one-year future temperature predictions in 2016, 2017, and 2018. For instance, model 2015 was trained with data until the end of 2015 and validated against 2016 data. Maximum temperature prediction r2 of 2017 South-West and North-East models were 0.96 and 0.98, respectively. Model 2017 has been deployed for alerting exception cases automatically and flagging abnormal temperature measurements. Also, the models improve the quality of heat injection design by providing temperature predictions based on planned heat injection rates. This novel automated workflow with data-driven models enhances reservoir management efficiency by reducing engineers’ unproductive time such as data manipulation and allowing them to focus on value-added works like analysis and optimization.

You can access this article if you purchase or spend a download.