Leveraging Machine Learning and Interactive Voice Interface for Automated Production Monitoring and Diagnostic

Abstract

Automated production monitoring and diagnostics is becoming essential for oil producers to achieve operational efficiency. In this work a combination of unsupervised and supervised machine-learning (ML) models are proposed and were integrated with interactive voice interface so that production diagnostic reports can be generated by using interactive session with chatbot. To achieve this, current work proposes an integration of ML models and chatbot in the cloud native environment and presents a case study using data from hundreds of wells supported on plunger lift system. Within ML framework data preprocessing and principle component analysis (PCA) was performed. The purpose of PCA was to identify principle components (PCs) and the projection production rate data over few dominating PCs and generate 2D or 3D plots which can be used to cluster wells based on production trends and relative performance. Then using daily production data, a regression tree analysis (per well) was performed to predict production rate for dominating phase for production. Regression tree generated if-else type rules which were used for production diagnostics. Further, using early few months of time series data for production, pressure and artificial lift data, another PCA model was trained and contribution chart (per well) were developed to identify which are the most contributing variables towards the change in the production such as increase or decrease in production rate. Finally, to enhance end user experience, a cloud native chatbot leveraging cloud services was configured to perform all steps involved in ML framework in serverless compute environment. The chatbot was built to answer frequently asked production monitoring and diagnostics questions such as "provide me a list of poor performing well" etc. The proposed framework was applied to wells supported on plunger lift and PCA revealed that that four PCs were enough to capture most dominating production modes and first 3 PC described 96.2% of variance. The diagnostic charts were built utilizing 2D and 3D diagrams using projection of gas production rate over first 3 PCs. This was found visually extremely useful to identify which well or group of wells were not performing as expected when compared to rest of the wells. Just by looking 2D plot about 10% wells were found with significant decrease while about 15% were found moderate decrease in production rate. Once identified poorly performing wells regression tree analysis was automatically generated along with the contribution charts for all variables. Couple of case studies were presented using two different wells with contrast production trend and it was demonstrated that the present workflow was able to identify relative behavior of those wells and presented detailed diagnostics using regression tree analysis and contribution charts. Overall, diagnostic charts were able to identify how to calibrate plunger count, plunger velocity, trip time etc. for improved production and forecasted up to 30% production improvement for poor producing wells. Finally, the results were tested out with chatbot. The chatbot model was deployed using web user interface and to answer production diagnostics related questions, chatbot utilized serverless compute to run ML models on the cloud. The output such as generated diagnostic charts and list of well etc. were prepared as user asked the questions and relevant analysis was presented to end user within a fraction of second. This can reduce time taken by well diagnostic analysis by 80%