Machine Learning Models to Predict Gas Hydrate Plugging Risks Using Flowloop and Field Data

Recently the concept of "no external gas hydrate control measures" has been proposed, whereby gas hydrate formation can occur in oil and gas subsea pipelines during steady state and transient operations, with the operational window defined by predictive analytic tools. Flow assurance engineers routinely use computer programs, including transient multiphase flow simulators coupled to a gas hydrate kinetics model to simulate gas hydrate formation and transportability. Given the complexity in multiphase flow modeling, modern machine learning technologies, especially artificial intelligence, could be applied to solve high-level, non-linear problems, such as evaluating gas hydrate risk based on measurable process parameters. In this work, several machine learning techniques, such as regression, classification, feature learning with an algorithm/framework like support vector machine (SVM) and neural networks (NN), are applied to analyze the data sets on: 1) hydrate tests conducted at pilot-scale flowloop facilities (4,500 data points), as well as 2) transient operation field data. The classification/regression model based on flowloop test data uses several independent input variables (features), such as water cut, gas-oil ratio, hydrate particle cohesive force, fluid velocity, oil viscosity, specific gravity, interfacial tension, and time in the hydrate stable zone, to output the hydrate fraction and probability of hydrate plugging in the pipeline. The semi-supervised learning model was applied based on the field data use as input, including water cut, shut-down time (where applicable), and gas-oil ratio to determine the level of hydrate resistance to flow during restart or dead oil displacement after production shut-down. The flowloop based machine learning model exhibited good prediction accuracies in test and validation processes, and was used to assess the hydrate risks in an actual field. The field data based machine learning model demonstrated the ability to construct field risk maps. The machine learning technique could be potentially applied in hydrate management to evaluate hydrate risks in subsea oil/gas pipelines. As a complement to more complex transient multiphase flow simulations, this machine learning approach can aid in the development of advanced hydrate management strategies.