Fracture Parameter Inversion in Geothermal Reservoir Using Deep Learning

Abstract Proper characterization of fractures is critical for evaluating the effectiveness of fracturing jobs and optimizing well performance in geothermal energy production, unconventional reservoirs, and other areas. However, accurately determining the size, shape, and orientation of these fractures solely from microseismic events is challenging due to weak signals and noise. To address this challenge, this study proposes a novel workflow that directly builds accurate fracture models from microseismic events using the DBSCAN clustering algorithm and BiLSTM-ESMDA. The first step is to filter the noise in microseismic events using the DBSCAN clustering algorithm. Next, a 3D planar equation is employed to construct the fracture plane in each perforation segment. Based on the results of this step, reservoir simulations are performed iteratively using PEBI grids and a BiLSTM surrogate model. Multiple representation models are obtained to capture calibration uncertainty and enable subsequent studies of long-term well performance, such as history matching for production. Finally, the ES-MDA history auto-fitting algorithm is utilized to find the most appropriate fracture model for matching production data through iterative processes. The developed inversion method was implemented on a representative geothermal model with a complex fracture network. The results demonstrate that the DBSCAN clustering algorithm effectively reduces noise in microseismic activity and ensures the accuracy of fracture geometry. A large number of different fracture models can be quickly generated by the surrogate model to capture calibration uncertainty. ES-MDA is utilized to optimize the fracture model and identify the optimal solution. The fracture models constructed using this method exhibit fracture half-lengths that are 20%-30% smaller than those estimated by microseismic monitoring. Furthermore, the high level of historical fit for this horizontal well indicates that the complex fracture model is realistic for the mine site. This study introduces a new approach to building a complex fracture network. By using microseismic data and BiLSTM-ESMDA, this method provides a practical solution. The proposed workflow significantly improves the accuracy of fracture network prediction and computational efficiency compared to traditional fracture inversion methods, which are often plagued by high multi-solution, high computational cost, and difficulties with convergence.