|Theme||Visible||Selectable||Appearance||Zoom Range (now: 0)|
Mracka, Igor (Slovak Academy of Sciences) | Zácik, Tibor (Slovak Academy of Sciences) | Hycko, Marek (Slovak Academy of Sciences) | Hajossy, Rudolf (Slovak Academy of Sciences) | Somora, Peter (Altova GmbH)
The purpose of the paper is to demonstrate the experience with distributed computing systems using evolutionary strategies for optimization problems in the gas industry. Two different approaches of distributed computing (cooperating agents and Client-Server) are presented.
Transient optimization of a gas transport is a long-term challenge , . Various optimization procedures have been proposed in the past, which excessively simplified the transport system: e.g., eliminated technological constraints on compressors and pipelines. In addition, a network without controllers or check valves is often assumed. However, these simplifications can lead to inapplicable solutions.
The aim of this article is to optimize the transition between two steady states using a sufficiently realistic simulation, with a complete optimization lasting no more than 30 min. It means to get sufficiently accurate results in a time sufficiently short for practice.
Our design is based on an already developed gas transport simulator (for 1 CPU) that takes into account a large number of real technological and contract constraints , . The optimization solution based on existing simulator opens a possibility of using other simulators in a similar way.
The article deals with reinforcement learning methods adapted to the gas transport management. Reinforcement learning, a part of machine learning, is nowadays one of the most active research areas in artificial intelligence. The goal of reinforcement learning is to learn good policies for sequential decision problems by optimizing a cumulative future reward signal. Notwithstanding of successes of reinforcement learning in playing games, its main practical role can be seen in exploring optimal decisions in managing real processes such as acting robots in an unknown environment, riding autonomous vehicles, and controlling industrial processes.
The purpose of the paper is to show how the reinforcement learning approaches can solve problems of a gas transport control in transient conditions. We will consider two situations. In first, an agent (controller) is learned to optimally and sequentially operate a real compressor station by adjusting its pressure set point. In second, it is learned to control revolutions of a single compressor (working area constrains are explicitly modeled). In both cases technical, operational and contract restrictions of pressure and flow are fulfilled. The agent must properly react when offtakes from the network change in time.
Learning process is simulation based and it utilizes an appropriate simulator of transient gas flow. Employed reinforcement learning methods must cope with continuity of space and actions. It is shown that, despite the long learning process, once the agent is trained, it is able to make right autonomous decisions, or at least help a dispatcher.
There are three main areas in machine learning: supervised learning, unsupervised learning and reinforcement learning. Supervised learning is based on training set or labeled examples provided by external supervisor. On the other hand, methods in unsupervised learning try to find patterns in collection of unlabeled data.
Reinforcement learning belongs somewhere between supervised and unsupervised learning. It depicts an autonomous agent who has interactions with surrounding environment and tries to learn how to choose optimal actions for achieving a desired goal.
Mracka, Igor (Mathematical Institute, Slovak Academy of Sciences) | Somora, Peter (Altova GmbH) | Hajossy, Rudolf (Mathematical Institute, Slovak Academy of Sciences) | Žácik, Tibor (Mathematical Institute, Slovak Academy of Sciences)
The article deals with a rupture localization system based on massive parallelized simulations on GPGPU (General-Purpose computing on Graphics Processing Units). The presented method is suitable for the localization of accidents on one pipeline and on a complex pipeline network as well. It can also be used in challenging situations where pressures at some inner network nodes are not available or in scenarios where emergency shutdown valves are closed and only one pressure sensor exists in a damaged pipeline section.
Main advantage of the suggested approach is the combination of highly precise simulation (to achieve localization accuracy) and a massive parallelization (to obtain the result fast enough). The described concept has been tested on both simulated rup-ture scenarios and data from a real rupture.
Fast and precise location of a rupture is important as it can prevent human casualties and environmental disasters after an accident. Prompt closing of shut-off valves minimizes the inevitable gas losses by stopping the delivery of gas to the rupture from the joined pipeline network.
There are many software-based leak detection and location methods but only a few for rupture location (Time-of-Flight Method based on the speed of rarefaction waves or a method based on a comparison of Real-Time Transient Model data with those from SCADA , ). A rupture location can be considered mathematically as an inverse problem to a pipeline simulation.
Highly nonlinear gas flow equations can be solved analytically but only approximately. When the first quick processes in a damaged pipeline just after a rupture disappears, the equations can be simplified to a heat transfer equation with known solutions. Using these solutions for pressure along damaged pipe-line in the rupture aftermath, the inverse problem can be solved analytically. It leads to four localization methods for specific cases (see ). As it was shown, only two methods for shut-off sections of a damaged pipeline are sufficiently precise and fast.
The aim of this article is to propose a solution of the inverse localization problem numerically. Presented approach is based on a comparison of real pressure measurements in the after-math of a real pipeline rupture with corresponding pressure outputs from simulations of ruptures at various positions. The location of the real rupture will be determined by the best agreement of the measured data and those data from simulations. The employed simulation has to be able precisely de-scribe high-dynamical pressure changes caused by a pipeline rupture.