Multi-fidelity machine learning with knowledge transfer enhances geothermal energy system design and optimization

Abstract

Designing and optimizing the control schemes of geothermal energy systems is a challenging and time-consuming work due to the vast parameter space and computationally intensive simulations. Canonical evolutionary optimization approaches are laborious, slow to converge, and may not provide optimal well-control scheme for geothermal energy systems. To tackle these issues, this work reports a machine learning-guided real-time flow control optimization for enhanced geothermal systems. This approach fully leverages existing data throughout the optimization phase by creating multi-fidelity surrogate models, which comprise coarse and fine models. The coarse model strategically selects a subset of variables to develop a low-fidelity representation, while the fine model utilizes all available variables to construct a high-fidelity surrogate. Knowledge transfer from coarse surrogate can guide the fine surrogate search into a promising subspace. Active learning technique is further leveraged to improve the accuracy of surrogate by iteratively querying the most informative data points. To evaluate effectiveness of the proposed approach, benchmark function suites and two fractured geothermal energy systems are employed in comparison with conventional evolutionary algorithms and advanced surrogate-assisted methods. The results illustrate the capability of the workflow to enhance the efficiency and effectiveness of real-time decision making. This workflow paves a new path for complex and computationally intensive design optimization problems.