Quantitative risk modeling of gas hydrate bedding using mechanistic, statistical, and artificial neural network frameworks

Abstract

Gas hydrates are crystalline compounds comprised of a network of hydrogen bonded water cages that can trap small gas molecules. These compounds are formed at the high pressure and low temperature conditions typically found in deep-water oil and gas pipelines. Gas hydrate formation in deep-water pipelines can lead to blockages, which can result in major environmental, safety, and economic hazards. This thesis focused on mitigating the formation of hydrate plugs in oil and gas pipelines. Specifically, the primary goal of this thesis work was to improve the understanding of hydrate bedding transitions by using three different approaches: mechanistic, statistical and machine learning (artificial neural networks, ANN). With the mechanistic approach, this thesis provided new insights into the important interconnection between partial water dispersion, agglomeration and hydrate bedding. This was achieved by developing a quantitative bedding framework that could consider water dispersion and agglomeration. A proposed model for the dispersion of water droplets in continuous oil phase systems included the prediction of the type and extent of water dispersion in a partially dispersed system; available literature models predicted the dispersion type (fully versus partially dispersed). In the statistical approach, 125 flowloop tests with approximately 5000 datapoints were analyzed and risk maps were generated. The analysis showed that linear regression models could be inadequate in predicting the hydrate plugging transitions in the flowloop. As an alternative, two initial Artificial Neural Network (ANN) based models (to account for nonlinear behavior of plugging transitions) were developed to quantify hydrate plugging risks. Mechanistic and ANN models developed during this thesis work could potentially aid in the development of an effective hydrate management strategy. Using the mechanistic approach, flowloop experiments performed using different oils indicated that hydrate agglomeration was intrinsically coupled to bedding. Additionally, large water droplets and partial water dispersion led to an early onset of bedding. Based on these two findings, a conceptual bedding framework was proposed by considering the effect of water dispersion and agglomeration. The new framework used a transient hydrate agglomeration model to generate a distribution of hydrate agglomerates as a function of droplet size distribution, particle cohesion force, shear and hydrate formation rate. The new mechanistic bedding framework predicted the onset of bedding with reasonable accuracy (coefficient of determination = 89%) for experiments performed at the ExxonMobil flowloop facility from 2014-16. Flowloop experiments showed that pressure drop due to fluid flow increased significantly after a certain hydrate concentration, defined as the plugging transition, which was considered a trigger point leading to hydrate bedding and plugging. Using the statistical approach, the flowloop hydrate plugging transitions were analyzed empirically using the pressure drop, particle size, mass flow rate and gamma-ray density measurements. A statistical approach was undertaken to determine the plugging transition from previous 125 flowloop tests, performed under various operating conditions. Data analyses suggested that the plugging transition could serve as a precursor to hydrate bedding. Tests with the King Ranch Condensate took significantly shorter times and lower hydrate fractions to reach the plugging transition criteria compared to the crude oil (with natural surfactants). Data analyses also suggested that low water cut, high mixture velocity, low Gas-Oil Ratio (GOR), and the injection of anti-agglomerants (AAs) can increase the plugging transition values and lead to safer hydrate transportation. It was observed that statistical linear regression models had limited accuracy in prdicting plugging transitions. In the third approach to overcome the limitations of the previous statistical approach for predicting the plugging transitions, two Artificial Neural Network (ANN) based models were proposed to quantify the plugging transitions in the subsea flowlines. The ANN based models were able to estimate the relative pressure drop and a binary class of failure to determine the “plug” or “no plug” case due to increasing hydrate concentrations in the flowloops. The ANN classification model also resulted in superior prediction accuracies as compared to the other classification models (Logistic Regression, Decision Tree, Support Vector Machine, and K-Nearest Neighbor). It was suggested that artificial neural network could be potentially useful as one of the tools for a hydrate management solution. During this thesis work, it was observed that the three approaches had their own unique advantages and disadvantages. The mechanistic risk model was able to explain the underlying physics involved in the dispersion, agglomeration (population balance model, PBM) and bedding processes. The mechanistic model was computationally intensive, as it required several numerical iterations while solving the differential equations with the agglomeration and breakage kernels. The statistical plugging risk model based on multiple linear regression had less complexity and was easy to interpret but gave poor predictability. The artificial neural network based risk models showed greater accuracy in predicting the plugging transitions, but had complex network architecture and were difficult to interpret. Nevertheless, given the complexity of hydrate particle transportability, these multiple approaches can contribute to an improved understanding and ultimate implementation of an effective hydrate management strategy.