Machine-learning-based well production prediction under geological and hydraulic fracture parameters uncertainty for unconventional shale gas reservoirs

Shale gas production prediction under history-matching-based geomodel is crucial to achieve reliable assessment and economic management of unconventional shale resources, however, conventional history matching is generally performed through repeatedly running high-fidelity reservoir simulations and therefore presents intensive computation-cost in practical applications. Without the need of history matching step, this work presents an efficient and robust post-history production forecasting framework using a latent-space learning-based direct forecast approach, e.g., referred to as LS-LDFA. A novel dimensionality reduction method, e.g., convolutional autoencoder, is employed to regularize the multi-well time-series data by low-order representation within a latent space. Once the machine-learning proxy is trained offline, the online post-history production forecast can be efficiently achieved by input history data. This paper presents some comparative studies between LS-LDFA and model-based history matching. This approach is tested on two examples with an increasing complexity, e.g., a multi-fractured horizontal well and a naturally fractured reservoir model with multi-well-pad-production based on synthetic shale formation. The results confirm that the method achieves high robustness and computational efficiency simultaneously in comparison with the conventional history matching. The application of learning-based direct forecast approach can effectively fuse information from history data and thus support reliable decision-making and risk assessment.