A Hybrid Approach for Production Forecasting and Uncertainty Characterization

RGNet (Reservoir Graph Network) is a pragmatic reduced-physics model that utilizes the essential physics of flow in porous media while simplifying computational complexities. Its versatility has been demonstrated through successful applications in a range of reservoir management applications, including reservoir connectivity analysis, resource volume estimation, dynamic forecasting, well control optimization, flood optimization, and integration with surface network models. While RGNet has been proven effective, there is a vision for further enhancing its modeling precision and predictive capabilities by incorporating advanced machine learning (ML) techniques. In this study, we propose a hybrid approach which combines RGNet and ML for production forecasting with uncertainty characterization. The basic idea is to use RGNet-simulated data as training features and use sparse regression (SpR) to select the important features and fit a data-driven model between the features and the target output. To forecast into the future, we substitute the simulated data of the future period into the SpR model to obtain the corresponding forecasted target output. For training purposes, we only need a rough prior model to start with and the calibrated model is not required, which significantly reduces the manual effort in history matching. In addition, as the training and prediction are very fast to perform, the same procedure can be repeated for multiple prior realizations in quite short time, which enables the uncertainty characterization of the future performance. The proposed hybrid RGNet-SpR method was tested on a deep-water field case to see whether it can improve the forecast performance of BHP. For training purposes, multiple prior RGNet realizations are randomly generated and simulated to generate the training features. The features considered relevant include simulated well-level injection/production rates and BHP data. As the simulation system are controlled using gas/water injection rates for gas/water injectors and using oil production rates for oil producers, these control variables as input features are also the historical production data in the training process. The output targets are BHP of target wells, which include multiple wells. As RGNet-SpR fits coefficients for different features, we use the magnitude of the features to characterize the strength of the connectivity between the target well and feature well. The proposed workflow is quite robust even when we use less than 50% of available data for training to obtain good quality forecasts. Also, multiple forecasts are shown to span the unseen data with a relatively small error band. Combining RGNet with ML provides a more robust and efficient solution for production forecasting with capability for uncertainty characterization. The proposed solution does not require a history-matched model as the base model, which saves resources and enables faster decision making.