Deep-Learning-Based Direct Forecast Approach for Fractured Shale Gas Reservoir Under Geological and Hydraulic Fracture Parameter Uncertainty

Abstract

Abstract Summary Accurate ultimate recovery prediction and uncertainty quantification are of significance during the process of unconventional shale resources development field development plan formulation. The uncertainties related to characterization of geological parameters and especially hydraulic/natural fracture system are not readily quantified through conventional model-based history matching procedure in practical applications. Data-space inversion (DSI) is a recently proposed data-space analysis and rapid forecast approach that easily samples the posterior predictions based on an ensemble of prior predictions and historical measurements. This paper presents a novel methodology based on a hybridization of DSI and a vector-to-vector residual neural network, referred to as DSI-DL, for well production forecast in unconventional shale reservoirs. A data-augmentation strategy that has been exploited in the literature is employed to generate a large number of training samples from a relatively small ensemble of shale reservoir model simulations. The DSI-DL method treats the shale gas production in our expected future time as target variables, which are directly predicted from pre-trained deep-learning model given historical data. This method enables us completely avoid the time-consuming history matching process where the characterization and inversion of hydraulic/natural fracture topology are highly complicated. We demonstrate the performances of this new approach on a synthetic vertically fractured well and a multi-well production scheme in naturally fractured shale formation. Some comparison with conventional DSI procedure and model-based history matching have shown that DSI-DL method achieves relatively robust results in estimating P10-P50-P90 values of shale gas time-series production against to data noise and prior ensemble size. DSI-DL algorithm will greatly contribute to the real-time prediction and optimization of shale gas well production rapidly given the streaming online observation data and have a wide range of practical application prospects, while a high computational efficiency remains. The proposed DSI-DL approach definitely will be useful for petroleum engineers to assess the value of information and manage the uncertainty of unconventional resources development.