Ultrafast Multiphase Production Forecasting for Large Gas Condensate Shale Reservoirs

Abstract

ABSTRACT Accurate production forecasting is essential for effective field development. Conventional techniques such as numerical simulations, reduced-order modeling, and decline curve analysis often require complex physics, detailed calibration, and significant parameter assignments, resulting in time-consuming and computationally challenging processes, especially for large, high-resolution geomodels. To address these limitations, we present a novel machine-learning-assisted modeling technique for rapid and accurate production forecasting. Our approach involves extreme geomodel compression, reducing the size by a factor of 20000, followed by machine-learning-based forecasting. We developed a multi-attribute, multi-layer shale geomodel compression technique utilizing dimensionality reduction. This method efficiently compresses large, heterogeneous shale geomodels into low-dimensional representations. Subsequently, a neural network model is trained to process the compressed representations along with completions and production parameters, enabling simultaneous prediction of monthly condensate and gas production rates over a 5-year decline period for hydraulically fractured shale wells. The forecasting workflow is evaluated on 3000 realizations of hydraulically fractured horizontal well placed in a complex heterogeneous shale reservoir. Each realization represents a condensate shale reservoir consisting of 88200 grid cells with spatially heterogeneous distributions of porosity, permeability, and connate water saturation. Our production forecasting method is evaluated using well-known metrics, including mean absolute error (MAE) normalized with respect to the target range (NMAE) and mean absolute percentage error (MAPE). Production forecast achieved an average NMAE and MAPE of 0.008 and 3.75%, respectively for the gas rate, in addition to NMAE and MAPE of 0.007 and 2.84% respectively for the condensate rate, confirming accurate predictions throughout the 5-year decline period. Furthermore, our workflow demonstrated significant efficiency gains, with forecast generation taking a mere 0.1 seconds after model training compared to over 10 minutes using commercial software. The model-building stage, including geomodel compression and hyperparameter tuning, required approximately one hour, depending on data complexity. Our methodology offers a valuable tool for economic evaluation, history matching, and production forecasting of both conventional and unconventional assets.