Understanding well production performance in hydrocarbon reservoirs in a timely manner is essential for closed loop reservoir management, improving operational efficiency, and maximizing value. It is desirable to have a robust and scalable method for estimating well productivity index and reservoir pressure, which can be applied in a practical and automated manner. Traditional surveillance methods are interpretive and do not scale for manual surveillance of either large fields or those with large data volumes. In this work, we propose a machine learning approach to discover physics that can be built using routine field measurements (downhole pressure and rates) and used for estimating well productivity, real-time production rates, pressure depletion and short-term forecasts.
The relationship between rates and pressure evolution is guided by nonlinear diffusivity equation. We seek methods for projecting the nonlinear state problem onto a linear (or weakly linear) space based on several methods – namely, time delay embedding, physics-inspired features, and dynamic mode decomposition (DMD). This augments the information contained in the system state with measurements of the state history. We also developed a background signal decomposition method to extrapolate routine buildup pressure data to estimate average reservoir pressure based on two different methods – optDMD (optimized DMD) and SINDy (sparse identification of nonlinear dynamics).
The background signal decomposition method was validated on several heterogeneous reservoir cases to estimate average reservoir pressure from buildup data, where our results outperformed traditional methods. In cases where multiphase flow meter rates were available, the proposed hybrid reservoir model was used to predict pressure with a virtual shut-in simulation. By offsetting the need for shutting in the well and associated production deferment, the virtual shut-in predictions were used to estimate reservoir properties. The results were validated on both pressures and pressure derivatives, typically used for pressure transient analysis. Next, we observed that the model can be used to provide accurate multiphase production rate forecasting (virtual metering) by reversing the model inputs and outputs. Based on the hybrid model, a workflow for tracking reservoir properties was developed to capture the decline of average reservoir pressure and productivity index, which was applied to both synthetic and field cases with reasonable accuracy.
The proposed hybrid reservoir modeling approach automates routine surveillance at field scale with high computational efficiency. By learning from natural operational variations continuously, it decreases planned downtime and associated production loss. It also enables detecting well performance issues much earlier to plan timely remedial actions. It provides a practical way of combining data-driven methods with our understanding of physics, while keeping the analysis interpretable and enabling closed loop reservoir management.
