Abstract:
The heterogeneous nature of many geothermal resources can be observed in existing geothermal operations when an anomalously low permeability well is drilled amidst high permeability wells. This study looked at a low permeability, weak production well that shared a two-phase pipeline header and separator vessel with several stronger production wells on the same well pad. The utilisation history of the well under investigation had shown episodes wherein the well stopped producing due to its low discharge pressure, even at throttled conditions. Such episodes usually lead to additional costs for the power plant operator as the well has to be stimulated to reinitiate its discharge. Real-time wellhead pressure data was used in this work to develop classification models that would reliably predict the occurrence of such low discharge pressure events in the future. A good performing model would allow the operator to perform intervening measures to prevent the complete collapse of the well, thereby minimising well downtime and avoiding unnecessary costs. A workflow based on systematic time-series feature engineering was applied to characterise the dynamics of the wellhead pressure trend. Machine learning models built using the relevant time-series features were able to reliably predict the occurrence of the low discharge pressure events in the target well. A numerical reservoir model was developed and calibrated to match the transient pressure response and production history of the target well. The model results showed significant fluctuations in the fractional dimension parameter before and after the low discharge pressure events.
