Application of Machine Learning Algorithms to Predict Plant Process Upsets During Well Cleanup

Abstract Tengizchevroil (TCO) has several stages of put-on production (POP) process to bring the new wells online. New wells need to be cleaned up before being fully brought online and directly routed to the plants. However, reservoir conditions in Tengiz require drilling and completion methods that often results in losing significant amounts of drilling and completion fluids which comprises of barite, cuttings, and emulsions. The presence of such drilling and completion materials may and does cause plant upsets which consequently leads to significant LPO (Lost Production Opportunity). To minimize plant performance upsets and equipment problems during a controlled well ramp-up of new, reworked, and stimulated wells (or, in short: plant flowback), a prospect of data analytics application arose intending to study the leading hitters that affect the processing at the plant during flowbacks and subsequently optimize the well flowback process. By applying machine learning and statistical methodologies, a machine was given the ability to perform prognosis on plant upsets and duration of the plant flowback. The approach is broken into three main stages. Stage 1: create a consolidated history of plant flowbacks. Stage 2: determine variables that have the highest impact on plant performance (this stage would reduce the complexity of the model by removing irrelevant variables and subsequently increase the accuracy of the machine learning model). Stage 3: build an algorithm that predicts plant upsets from the variables extracted in Stage 2 (provides a data-driven method for evaluating risks associated with plant flowback). A supervised classification model was built on historic data from 2011 that evaluated plant performance risks associated with the well flowback process. Given the risk probabilities, we could predict how long it would take to complete plant flowbacks on a subset that the model had not seen before. Additionally, it was shown that given the evaluated risk for the subset the model has not seen, a sequence in which these wells were put online could have been optimized to maximize production. The feasibility of machine learning capabilities was tested on the historic plant flowback data. The study results have confirmed the intuition of the subject matter experts but in a robust and data-driven way. The model and the approach show data analytics methodologies’ applicability to optimize production operations further.