Deep learning-based inversion framework by assimilating hydrogeological and geophysical data for an enhanced geothermal system characterization and thermal performance prediction

Enhanced geothermal systems (EGS), which are developed by creating artificial fractures to enhance the permeability of deep reservoir, show their its advantage in power generation. However, due to the limited wells in deep depth, characterizing fractured geothermal reservoirs with sparse data is difficult and poses challenges for long-term thermal performance prediction. Interpreting observation data through inversion to characterize the fracture aperture and predict EGS long-term thermal performance is a hopeful scheme. Thus, a joint inversion framework, convolutional variational autoencoder-ensemble smoother with multiple data assimilation (CVAE-ESMDA), is proposed with the following aspects: (i) CVAE is trained to parameterize the high-dimensional Non-Gaussian aperture field, (ii) ESMDA is applied to integrate hydrogeological and geophysical data for aperture characterization. Based on the inversion results, the long-term performance is predicted. The normalized root mean square errors (NRMSEs) of inversion results decrease from 27.66 % to 24.52 % after multi-type data are integrated for CVAE-ESMDA. Furthermore, the NRMSEs of long-term thermal prediction of two production wells also decrease to only 2.3 % and 8.2 %. To further exhibit the advantages of CVAE-ESMDA, another joint inversion framework, principal component analysis (PCA)-ESMDA, is also compared. However, PCA is limited in addressing Non-Gaussian aperture field and its long-term thermal prediction performance is unsatisfactory.