Preventing Slugging by Tuning Choke through Machine Learning

Abstract

A gas-lift well sometimes suffers from slugging. As slugs reduce production volumes and cause other issues on the surface, we would like to mitigate or avoid them. The production choke and gas injection choke are two points at which the operator may influence the slug. For this to work, the operator must know that a slug is going to occur in advance so that avoidance actions can be implemented. The operator also needs to know by how much to change each choke. We find that a slug can be forecast successfully five hours in advance given typical field instrumentation of the well. This is based on an LSTM machine learning approach given historical data only.