ESP Failure Prediction in Water Supply Wells Using Unsupervised Learning

Abstract

Water injection is seen as one of the key field development strategies to achieve the mandated production target as it will maintain reservoir pressure as well as improve sweep efficiency and increase field recovery factor. In current practices water supply wells workovers are planned after Electrical Submersible Pumps (ESP) are failed by adopting run to fail approach. This lead to decrease in well availability and increase in down time which impacts water injection cluster capacity in giant matured onshore oil field. The objective of this solution is to early detect the failures for ESP wells using Machine Learning (ML), by demonstrating the feasibility of this approach and verifying that the concept has practical potential, the tool can be used to reduce deferment and increase well availability either by extending time-to-failure or better planning and scheduling the workovers. In this solution, Predictive Analytics model was developed based on Algorithms using field sensor data, and well failure history to predict ESP well failure probability. Due to the limited available ESP real time data, it would be a challenge to have an accurate model. The downhole and temperature data is not available in these ESP wells. Hence, we have adopted unsupervised classification approach combined with statistical calculations such as MTBF based on failure history. The solution provides a probability of ESP failures based on the anomalies (anomaly severity) detected from unsupervised machine learning model (individual cluster based), MTBF & number of starts. The probability is normalized based weight-based approach. Additional criteria can be added and considered in the future to fine tune the model and predictions. The approach has successfully evaluated on 34 water injection clusters in this giant field. The model is able to predict 77% of failures historical failures successfully. The limitations in ESP down-hole data availability and real time quality issues impacted model accuracy. The solution has been successfully deployed in real time mode and able to predict failures 90 to 120 days before failures. This has resulted increase in well availability by 10% and increased water injection system capacity. This machine learning based approach has been extended to all water injection clusters and also capitalized in other fields to increase well availability and grow capacity with the increasing demand for water injection to sustain and grow production volumes