Automating Dynamometer Charts Interpretation with Machine Learning

Del Pino Fiorillo M.A. · SPE Latin American and Caribbean Petroleum Engineering Conference Proceedings (2023) · DOI: 10.2118/213191-ms

Resumen

This work objective is to develop a comprehensive methodology for constructing a mathematical modelcapable of automating the interpretation of Dynamometer Charts from Sucker Rod Pumps (SRPs).The methodology describes the complete data-handling pipeline that allows for the proper ingestion,normalization, transformation, and evaluation of charts, comparing the predictive performance of multiplewell-known Machine Learning models. The methodology delivers a properly tuned model with a highpredictive performance that can interpret any chart shown to it, as well as quantifying the similarity of anygiven chart to that of pre-defined behavioral classes. To develop the methodology, a sample of 293 Dynamometer Charts was evaluated to determine theequipment status, based on known patterns described in the existing bibliography. The resulting datasetwas used to train and test around 1,000 different models. These models include Support Vector Machines,Decision Trees, Nearest K-Neighbor, Neural Networks, among others. Each model was tested multiple timesto account for bias and overfitting. This way, a best-performing model was found based on performancescores, allowing to determine model family and hyperparameters required for optimal performance. The study demonstrated that many Machine Learning models can be properly trained for performinginterpretations of SRPs operational status. The best performing models were those based on the XGBoostalgorithm, although Random Forest algorithms proved to be significantly accurate in certain circumstances.Results were differentiated between surface dynamometer and bottom-hole dynamometers, with theintention of demonstrating the potential implementation of models with no data manipulation in between,such as the one caused by using the wave-function transformation. This paper not only shows how Machine Learning algorithms can be used to automate some workflowsof the production engineering discipline, but also provides a methodology for validating model performanceand determining the best performing model among many candidates. Finally, it shows practical ways inwhich these models results can be used to generate visual tools to improve the decision-making process ofengineers and operators.