Abstract:
The time consumed between observing the production decline and identifying the liquid loading problem is a big challenge, facing the production engineer. The objective of this project is to predict the liquid loading condition using a machine learning approach. This work included visualizing, preprocessing and modeling the data by k-nearest neighbors regression algorithm. For this study gaseous wells were selected. Production and completion history for each well were collected. First study was performed on (synthetic) data where 70 percent of the information were used for the training purpose, 15 percent for calibration and 15 percent for validation of the model. The model successfully anticipated the liquid loading status with an accuracy of 93% and 93% of the data trained. A local data obtained from the well (FN 21) had experienced an attempt to be modeled by the same way, but due to lack in completion and production data, the attempt has failed. Another local data obtained from the well (FN4-7) with complete completion and production data was modeled. The model successfully predicted liquid loading status with an accuracy of 92 % and 100% of the data trained. This new smart model developed for local data shows a great promise that this approach can be applied in other areas where a limited history of production and liquid are available.
