Large Pipeline Network Optimization: Summary and Conclusions of TransCanada Research Effort

Abstract

Operation of large gas pipeline networks calls for fulfilling variation in contractual volume obligations, and maintaining a certain range of linepack with minimum fuel consumptions to drive compressor units. This is often achieved with either operational experience or by utilization of optimization tools, which results in reduced hydraulic analysis time as well as improved pipeline operation as a whole. The main objective is to accurately identify the optimum set points for all compressor stations, control and block valves in the network, subject to several system and operational constraints. This implies multi-objective optimization of a highly constrained network with a large number of decision variables. Over the past three years, TransCanada has devoted a research effort in developing/integrating an optimization tool based on stochastic methods. It was found that it offers greater stability and is more suited for multi-objective optimizations of large networks with inherently large number of decision variables, than any gradient-based method. This paper describes the nature of the pipeline system under optimization, and discusses the basis for a Genetic-Algorithm-based tool employed. It summarizes the results of the past three years of research efforts outlining the selection criteria for the optimization parameters, integration with a robust steady-state thermal hydraulic simulator of the pipeline network and the notion that dynamic penalty parameters can affect convergence. The methodology is applied to a large gas pipeline network containing 22 compressor stations resulting in 54 decision variables and an optimization space of 1.85×1078 cases. Comparison of genetic algorithm optimization with traditional and manual optimization is demonstrated. Extensive effort has been devoted to reduce the computation time, which includes techniques to utilize various hybrid surrogate methods such as Kriging, Neural Networks, Response Surface, as well as exploitation of parallel processing.