Accumulation of sand within pipelines can pose significant problems in the oil and gas industry. Sand particles tend to settle out of suspension leading to the formation of stationary or moving beds along the bottom of pipelines. Such beds provide an environment that result in flow assurance problems such as increased corrosion rates, increased pressure drop in pipes and pigging blockage. Thus, the prediction of sand transport velocities is of great importance for flow assurance in petroleum as well as mining industries. The processes in these applications must be designed and operated at a sufficient fluid velocity to avoid solid deposition. Recent studies at the Tulsa University Sand Management Projects (TUSMP) have shown that Artificial Intelligence – Machine Learning (ML) methods can be effectively and accurately used in predicting minimum particle transport velocities in pipelines. However, these methods have not been rigorously developed and tested for multiphase air-water flows with particles. The purpose of this work is to investigate the use of several ML models to predict the critical velocities in horizontal and inclined multiphase flow pipelines.
In this study, three machine learning algorithms, including Support Vector Machine, Random Forest, and Extreme Gradient Boosting, are utilized to predict minimum flow rates required to transport particles successfully in intermittent and stratified gas-liquid flow regimes. The models predict the value of critical velocities in pipes via ML using accessible parameters as inputs, namely, sand concentration, pipe inclination, pipe size, liquid density, liquid viscosity, particle density, and particle size. First, these models are trained with a set of 1640 data points. After hyper-parameters of each model are optimized, the results are verified with a test data set and their predictive abilities are cross-compared. After the final set of models is constructed, an error analysis was performed by evaluating the results when input parameters, such as superficial velocities and fluid properties, were changed. Later, the predictive performance of the method is also validated using out-of-sample data available from the literature. Finally, the predictive abilities of the best model are further validated by comparing its performance with well-established mechanistic models based on empirical correlations. The Random Forest results reveal a better training performance and prediction. The results also indicate that the proposed method gives comparable or even higher scores by contrast to correlations and mechanistic models for multiphase flow, and could be easily employed for industrial applications. The application of the above-mentioned ML algorithms and the large database used for their training allowed extending the proposed methodology to a wider applicability range of input parameters as compared to standard accessible techniques. The ML results present competitive and even more accurate predictions as compared with the existing mechanistic models, indicating a great potential of utilizing the data-driven machine learning methodology for applications in flow assurance.