Pressure Drop in a Network of Pipelines Using Machine Learning Models

Abstract

Field development planning aims to produce hydrocarbon accumulations while minimizing costs and maximizing revenues. Facility placement and pipeline layout development and optimization plays an important role in a field development planning process. Pressure drop is a significant parameter used to evaluate the different scenarios for well and facility placement optimization. Optimization employs a large number of simulations each including pressure drop estimation through a production network. Therefore, the process of estimating the pressure drop through the pipelines takes place numerous times and hence the need of an efficient algorithm that automatically builds and solves the surface network. To address this challenge, we proposed a modular approach to build machine learning models with a real world application using synthetic data generated from PIPESIM simulator. The modular approach allows us to build various ML models with different features and ranges to address different cases. Then we integrated the developed ML models into a network solver to solve for pressure drop through a single pipeline or a network of pipelines. The ML network solver was built in a flexible way where ML models can be easily added as much as required. The ML network solver was able to solve a set of networks generated by OGED Field development Planning interface. We have developed eight ML models to cover the length and inclination ranges in the networks. The results for pressure drop, component flowrates, and pressure at inlet nodes of the pipelines were promising where the maximum error among all the networks was 8.79%. However, the processing time was not efficient compared to PIPESIM simulator run in a batch mode. Although PIPESIM simulator performed better in terms of processing time compared to the ML network solver, the PIPESIM simulator processing time remained high. To accelerate the process of modelling and solving a pipeline or a network, we introduced upscaling. The upscaling algorithm was applied on the same networks addressed earlier. The processing time decreased when using both the PIPESIM simulator and the ML Network Solver. The ML network solver results were also promising where the maximum error was 10.88% although its speed remained inferior to PIPESIM simulator. Machine learning techniques are recently used to predict pressure drop in pipelines only. However, to the best of our knowledge, they are not yet addressed in such a modular approach and applied for pressure drop evaluation through networks. Our approach is an innovative way to address different real world scenarios. Once successfully integrated with optimization algorithms for facility placement optimization, the proposed work will be a valuable addition to the oil and gas industry.