Traditionally, pipe flow models have been derived from engineering data (such as pipe and fluid properties) by constructing the appropriate nonlinear partial differential equations . This engineering data comes from a variety of sources and is often wrong. Therefore it has been necessary to tune these models, either manually or automatically, until their results matched the real-world behavior of the pipeline. However, with the aid of some usually-valid assumptions about the underlying equations, it is possible to construct a transient pipe model solely from historical SCADA data, thus eliminating the need to find, input, or validate pipe or fluid property information. This paper presents a method for automatically generating such models, studies the accuracy of the generated models for a real gas pipeline, and examines some possible applications.
Online models are one of the most useful applications made by the pipeline simulation industry. They have many different functions including detection of leaks and other anomalies like failing equipment, describing what is happening in the line at unintsrumented locations, and maintaining an accurate picture of the current state of a pipeline for use in launching lookahead models. Look-ahead models, which are close relatives of online models in that they depend on SCADA data to give them their initial state, can determine feasibility of planned operations, survival times, and ETAs of important events. However, creating an accurate online model can be difficult. Even assuming that the underlying model engine is correct and bug-free, configuring these programs depends on getting an accurate description of the physical properties and layout of the pipeline itself, the properties of the fluid being pumped, and numerous other numbers that must be drawn from the real world. There is plenty of opportunity for human error while entering this data into the model; in addition, much of this information may not have been used for ages, so the pipeline operator would not necessarily know if it was still accurate. On top of this, the properties of pipelines change over time. This necessitates some sort of automatic adjustment of the original properties, which is typically accomplished by "tuning" pipe roughnesses, fluid properties, ground thermal conductivity, and the like. This is especially necessary in the case of an application like leak detection that requires high model accuracy. This sounds like a lot of work, and it is. However, there's an alternative to building a model starting from a physical description of the pipeline: you can instead start from the observed behavior of the pipeline. If a program can accuately predict what some of the SCADA instrumentation will read given values for some of the other SCADA points (e.g. predict the flows if it knows the pressures, or vice versa) then it will effectively be a model, capable of fulfilling a subset of the online and look-ahead model applications described above.