Spe-192002-ms case Study Applied Machine Learning to Optimise pcp completion Design in a cbm field
Censored data and multiple failure modes
Download 1.89 Mb. View original pdf
|
- Navigate this page:
- Data Exploration and Feature Importance
- Table Example of Model Dataset
| Censored data and multiple failure modes The residuals of the model run with all the production periods even the ones where the pumps have not failed proved to have a higher mean absolute error of 154 days when compared to the model with only failed production periods (126 days. There is a heteroscedacity introduced by the inclusion of production periods where pumps have not failed causing the model to underpredict pump runlifes. The reduction in heteroskedacity comes with a trade-off in that the number of observations also had to be reduced from 1499 to 895. This needs to be a consideration when building a recommender system. The presence of censored data and multiple failure modes potentially compromises the validity of the likelihood calculation used to obtain the posterior. Further work is required to understand the effect of censored data and multiple failure modes on model performance. Data Exploration and Feature Importance During the data exploration phase, 137 features were created from the available data. Of those only were kept for the final model, these features can be seen in table 2 . The key features were well parameters (top-depth perforated interval, pseudo steady water rate, cumulative water rate) and completion parameters (pump geometry and displacement, Rod OD and Tubing ID). Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/18APOG/2-18APOG/D021S016R002/1220497/spe-192002-ms.pdf/1 by Vedanta Limited - Cairn Oil & Gas user on 28 June 2023 SPE-192002-MS 7 Table Example of Model Dataset The difference in figure 2 and figure illustrate the reduction of entropy or noise that the important features can add to a model. The model that generated figure used only pump design features with all subsurface features removed. Grid Search Optimisation Once the features and the model were decided upon a simple multivariable grid was used to iterate through the different pump design choices fixing the well representation. Here the power of the gaussian process regression comes into its own. Not only can the model generate a prediction it also generates the variance or confidence value around its prediction. The recommender system can be used to either exploit the design choices with good predicted runlife with the lowest variance (the conservative choice) or explore the design choices with higher predicted runlife and a higher variance (an experimental choice. The last option is worth exploring to explore pump designs options that may potentially lead to higher runlives. The runlife surface being explored is illustrated in figure Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/18APOG/2-18APOG/D021S016R002/1220497/spe-192002-ms.pdf/1 by Vedanta Limited - Cairn Oil & Gas user on 28 June 2023 |