Predicting Shale Oil Well Production by Transfer Learning from Large-Scale Field Data

Summary Accurate prediction of shale oil production is crucial for field designs. Though various methods have been proposed in the literature, it is challenging to predict long-term production with limited early history. In this work, we use dynamic time warping (DTW), long short-term memory (LSTM) neural network, and transfer learning theory along with the production data from more than 30,000 shale oil wells in North American oil fields and the Changqing shale oil field in China to improve the production prediction accuracy of new wells. The following results are obtained: First, the DTW method is effective in identifying similar wells, where smaller DTW distances correlate with higher similarity. The DTW distance ranges from 0.02 to 0.13 within the Eagle Ford (USA), while the DTW distance between Eagle Ford and Bakken (USA) wells becomes 0.05–0.35. The Changqing shale oil well is more similar to Bakken shale, with DTW distances of 0.25–0.75. In the three major shale oil-producing areas in North America, for wells in their early production stage, we can successfully find wells with high similarity in the database that have been produced for a long time. The proposed prediction workflow combining LSTM and transfer learning performs best compared with directly using similar well data, traditional decline curve models, and a single LSTM model. The prediction accuracy of the LSTM and Arps models improves with increasing production time. When applying the large North American data set to predict the performance of shale oil wells in China, the proposed method improves prediction accuracy by 22.6% compared with using only matched well history and by 50.09% relative to the selected decline curve model. This work not only introduces an effective way to use the extensive existing shale oilwell data but also demonstrates the degree of similarity in production curves across wells, basins, and countries.