Reliable Production Optimization Using Hybrid Physics-Based Models and Data-Driven Techniques

Abstract Production optimization of unconventional wells has great economic potential but is challenging due to issues related to high well count, complex reservoir conditions and production methods, and availability and quality of data, among others. A physics-based model is useful for predicting well performance required for optimization, since measurement data are often noisy and do not provide sufficient coverage of the entire parameter space. On the other hand, data-driven models are easier to build and significantly faster to run. This paper proposes a new hybrid methodology which combines a physics-based modeling approach with data-driven techniques for the purpose of optimizing production in unconventional assets. This paper presents a new hybrid approach utilizing both physics-based models and data to predict well performance. Mathematical models that simulate inflow, tubing hydraulics, and chokes are integrated to generate the prediction. To handle high well count efficiently, flow correlations are converted into a set of multi-dimensional tables covering broad ranges of parameter values and flow conditions. Measurement data of pressure and production rates are used to obtain optimal model parameters in those tables. The tuning process utilizes time series data with both pressure and rate measurements as well as time series data with pressure measurement only. New techniques have been introduced to effectively assimilate data of different types, which are key to making the hybrid method efficient and reliable. The proposed method has been applied to hundreds of wells in Permian for predicting oil, gas, and water production rates as functions of choke position and gas lift injection rate. Careful selection of the weighting scheme enables a close match between the model and measurement. The computational workflow generates predictions of optimal gas lift injection rate corresponding to each choke position. The resulting well performance surfaces are fed to control systems that maximize oil production while honoring different operational constraints. Initial deployment of this technology has resulted in oil production uplifts at several tank batteries. In comparison to the traditional physics-based modeling approach, the current method improves prediction reliability by automatically detecting gaps in the training data and borrowing data from analog wells if necessary. Different techniques have been implemented for effective removal of outliers in the measurements. The ability to assimilate data incorporating time series of different types further strengthens the predictive capability of the hybrid method.