Data-Driven Diagnosis for Artificial Lift Pump's Failures
- Thanawit Ounsakul (PTTEP) | Ake Rittirong (PTTEP) | Thanawee Kreethapon (PTTEP) | Wararit Toempromraj (PTTEP) | Kittipat Wejwittayaklung (PTTEP) | Phattarakorn Rangsriwong (PTTEP)
- Document ID
- Society of Petroleum Engineers
- SPE/IATMI Asia Pacific Oil & Gas Conference and Exhibition, 29-31 October, Bali, Indonesia
- Publication Date
- Document Type
- Conference Paper
- 2019. Society of Petroleum Engineers
- Optimisation, Artificial Intelligence, Data Sciences, Neural Network, Artificial Lift
- 1 in the last 30 days
- 161 since 2007
- Show more detail
- View rights & permissions
|SPE Member Price:||USD 5.00|
|SPE Non-Member Price:||USD 28.00|
Beam pump and ESP are common artificial lift techniques in pumping systems. They are widely used as primary oil recovery methods, but system failures lead to production deferments and increases in operating expenses. Employing decades of our field data, promising data science techniques are discussed here to analyze the factors governing failures in both beam pump and ESP approaches. These data are then applied with machine learning models to predict service life, failure mechanism performance, and production deferments.
The data analytics process begins with data preparation. Field data were extracted, transformed and loaded into a data warehouse for further processing. These data were categorized by failure information, pump configuration, wellbore geometry, and production information. The significance of each parameter causing pump failures was derived using a process called "Attribute Forward Selection (AFS)." Then several machine learning algorithms were implemented and compared with to determine the most appropriate model to predict pump service life. More suitable pump configurations to improve pump service life were conceptually recommended based on the analysis.
Differences in parameter significance was identified by attribute forward selection, and is displayed in a heat map. It was seen that the use of beam pumps in highly tortuous wells received the number one ranking as the main cause of failures whereas sand production was revealed as the most significant parameter relating to ESP failures. Correlations for these parameters were mapped by machine learning algorithms, resulting in multivariate failure prediction models (i.e. involving more than one parameter at a time) to predict the service life of beam pump and ESP systems. For both artificial lift systems, the models with the best correlation found thus far are based on a neural network, which resulted in the highest R-squared values when compared to other techniques. This neural network model was validated with the actual information, and the outcomes using this model are presented via a scatter plot in this paper. The plot shows that the prediction for ESP forms a trend around the theoretical best match line. In contrast, the prediction for Beam pump still needs improvement, with the data being scattered around the straight line with a unity slope.
Data science is an emerging technology that recently has provided breakthrough results for big data analysis. This paper will demonstrate the application of such discipline to the area of artificial lift. Machine learning is a promising tool which could help improve human understanding of complex problems, and, in this case, could furnish a durable competitive advantage to the oil and gas industry.
|File Size||867 KB||Number of Pages||10|
McDonnell C., "Towards Data Science," 3 February 2018. [Online]. Available: https://towardsdatascience.com/what-really-drives-higher-life-expectancy-e1c1ec22f6e1.