Abstract
This research presents a new architecture and implementation to overcome inherent challenges in leveraging machine learning (ML) for evaluating production performance in unconventional wells. By implementing a hybrid data-physics architecture (HDP), our goal is to effectively address several persistent hurdles, including generalizability limitations across diverse samples, the necessity for extensive training datasets, and the discrepancies between model predictions and fundamental physical principles. These fundamental constraints form the focal points of our comprehensive investigation.
This new HDP architecture seamlessly integrates a physics-based equation into the framework of a deep neural network model. The training dataset encompasses a wide array of influencing factors on production rates, encompassing information that may not readily conform to conventional physical equations. This sophisticated approach enables the inclusion of supplementary data, thereby significantly enhancing the precision of production forecasts. As a result, these data points are adeptly employed to derive the model's underlying physical parameters, leading to highly accurate production rate calculations. Once these parameters are estimated with minimal error, the trained model exhibits exceptional proficiency in forecasting both short-term and long-term production rates consistently.
To thoroughly evaluate the efficacy of the developed architecture, an extensive assessment was conducted using unconventional wells situated in the Duvernay resource within the western Canadian sedimentary basin (WCSB). This evaluation spanned three different methodologies to compute future production rates: with physical decline curve equations including Arps, Power Law, and Duong, with data-driven modeling where production rate forecasted with different powerful ML techniques including Random Forest, Ada-Boost, and K-Nearest Neighbors, and finally with HDP modeling. The results compared with different statistical metrics across all evaluated scenarios, the hybrid model consistently exhibited superior precision in production forecasting.
A noteworthy advantage intrinsic to the new hybrid architecture is its remarkable capacity to generate more accurate predictions without requiring extensive sample points for training. This characteristic proves especially advantageous for newly established wells with limited production histories. Moreover, the predictive outcomes yielded by the hybrid model demonstrate a strong alignment with fundamental physical models, thereby validating its applicability across both short-term and long-term production forecasting contexts.
