Traditionally, subsurface models are created based on reservoir characterization, then simulated and calibrated via history matching (HM) to honor data, generate forecasts, and quantify uncertainties. However, this approach is time consuming for unconventional projects with aggressive schedules. On the other hand, purely data-driven approaches such as decline curve analysis (DCA) are fast but not reliable for yet-to-be-observed flow regimes, e.g., boundaries or other effects causing late-time changes in productivity decline behaviors. We propose a physics-informed unconventional forecasting (PIUF) framework that combines simulations and data analytics for robust field applications. We apply Data-Space Inversion (DSI) to incorporate physics from a large ensemble of prior simulation models to generate posterior forecasts within a Bayesian paradigm. We also quantify the consistency of simulated physics and observed data by computing the Mahalanobis distance to ensure that the appropriate prior ensemble is employed. In lieu of history-matched models, a statistical relationship between data and forecast is learned; then posterior sampling is applied for data assimilation and direct forecasting in DSI. DSI reduces the dimensions of time-series (and other) data using parameterization like Principal Component Analysis. We implemented DSI within a tool that is connected to a vast database of observations for thousands of unconventional Permian Basin wells and a large ensemble of fracture simulations. We apply it to rapidly generate probabilistic forecasts (e.g., oil production rate, gas oil ratio) for unconventional wells and show that DSI can provide robust long-term forecasts based on early-time data when compared with DCA. We show that DSI yields robust uncertainty quantification with a manageable number of simulations compared with simple machine-learning methods like K-Nearest-Neighbors. We illustrate how data error and volume impact DSI forecasts in meaningful ways. We also introduce a DSI enhancement to generate posterior distributions for model parameters (e.g., hydraulic fracture height) to derive subsurface insights from data and understand key performance drivers. Our cloud-native implementation stores data (observed and simulated) in the cloud while the algorithm is implemented as a microservice that is efficient and elastic for the analysis of many wells. The overall framework is useful for rapid probabilistic forecasting to support development planning and de-risk new areas as an alternative to DCA or HM.

You can access this article if you purchase or spend a download.