Machine-Learning-Based Proxy Modelling for Geothermal Field Development Optimisation

Abstract

More than 40 billion tonnes of CO2 are released annually, hampering climate change efforts. The goal of current research is to utilise these gases in generating energy. The oil and gas industry faces increasing expectations to clarify the implications of energy transitions for their operations and business models, reduce greenhouse gas emissions, and achieve the Paris Agreement and Glasgow Climate Pact targets. A solution is integrating machine learning and geothermal energy to optimise field development to reduce CO2 emissions while meeting energy demands. The study area is a simulated actual field data, with three existing geothermal doublets and six exploration wells. The development plan aims to satisfy the energy demand for two locations, D1 and D2, for the next 100 years, using geothermal energy and optimising field development plans via machine learning models as surrogate models. A pseudo-geological model was developed using limited field data to identify sweet spots for further drilling. Four separate model cases were simulated using DARTS. The time-energy data from DARTS was then used to train and test several machine learning models to serve as a proxy model to optimise the best strategy to meet the energy demand. The economic model was simulated for 20 years for the selected strategy for field development. Using an injection rate of 500 m3/day per well to validate the ML models, the best-performing model had a mean absolute error within the range of 0.6 to 1.5 MW for all the doublets. Based on the ML results, the computational power and time required for field development plan simulation were dramatically reduced, and several configurations were performed. The optimal strategy for this field comprises 7 geothermal doublets, 3 for D1 and 4 for D2. This strategy uses all available wells to avoid lost investment or excess cost when those wells are needed to complement production when decline sets in after 20 years, allowing a reliable and long-term energy supply. This strategy will achieve a net energy output of 108 MW for D2 and 82 for D1. This strategy uses machine learning energy estimation for the optimum configuration and addresses the issues of excess energy storage, uncertainty in production, and rising energy demand. The economic model was based on a fixed OPEX, an estimated Capex based on field development strategy, and an associated discount rate of 7%. The project resulted in a Levelized Cost of Energy of €11.16/MWH for 20 years whiles reducing annual CO2 emissions by about 367,000 metric tons. This study shows that geothermal energy is a crucial step toward cleaner energy. ML can speed up the energy transition by optimising geothermal field development. This research aims to reduce CO2 emissions while meeting energy needs.