Rapid Production Forecasting for Hydraulically Fractured Wells in Heterogeneous Shales

Abstract

Abstract Production forecasting is crucial for field development. Machine Learning models have shown the potential to overcome limitations of the conventional methods like Numerical simulations, reduced-order modeling, and decline curve analysis. The paper presents a newly developed machine-learning-assisted rapid production forecasting method, involving massive geomodel compression (18000 times) followed by neural-network-based regression. The method first compresses the large, heterogeneous shale geomodel to a low-dimensional representation. Then, a neural network processes the low-dimensional representation along with completions and production parameters to predict the condensate and gas production rates of a hydraulically fractured shale well for a period of 5 years. In our study, the forecasting model is trained and tested on a heterogeneous dataset containing 3000 distinct realizations. Each realization is a condensate shale reservoir comprising 88,200 grid cells with spatially heterogeneous distribution of porosity, permeability, and connate water saturation. Subsequently, the generalization capability of the forecasting model is evaluated on a holdout dataset. Furthermore, we explored the impact of adding the early-time production history, spanning two to six months as additional input features to the forecasting model. Our objective was to assess how this inclusion might potentially reduce the required size of the training dataset while still achieving satisfactory forecasting performance. The performance of our newly developed production forecasting method was assessed using the well-established metric, the mean absolute percentage error (MAPE). The forecasting model achieved a MAPE of 3.2% for gas rate and 3% for condensate rate on the testing dataset. As for its performance on the holdout testing set, the model achieved a MAPE of 5.8% for gas rate and 4.4% for condensate rate. Furthermore, our study demonstrates that incorporating early-time production data, especially spanning over six months, significantly enhances the predictive accuracy, especially in scenarios with limited training instances. Inclusion of even a two-month production history provides substantial benefits compared to models devoid of early-time production data. When an early-time production data is included in the forecasting model, there is a significant reduction in computational time of up to 85% in generating the training datasets. Our research introduces a distinctive approach that combines advanced geomodel compression techniques with a multi-layer neural network model for production forecasting. The newly developed geomodel compression followed by neural-network-based rapid production forecasting reduces the computation time to generate a 5-year forecast per realization by an order of 6. Overall, the newly developed rapid production forecasting per realization takes approximately 0.0003 seconds as compared to 1037 secs or 17 minutes per realization for a traditional simulator.