TY - THES

T1 - Application of neural networks for short-term prediction of flowing wellhead pressures in horizontal gas storage wells

AU - Sachsenhofer, Markus

N1 - embargoed until 15-11-2021

PY - 2016

Y1 - 2016

N2 - The underground storage of natural gas and the security of supply thereof is an important issue for a country like Austria, which depends on imports of fossil fuels. Typically the operation mode of an underground gas storage facility is determined by the seasonal swing in the demand of gas opposed by the nearly constant supply through pipelines. A more short-term oriented operation schedule supported by a numerical simulation of various scenarios could improve the energy efficiency and reduce wear through optimized switching cycles. This master's thesis evaluates the application of neural networks to the simulation of gas storage behaviour. Neural networks are heuristic tools that can be used to model nearly any trend in the data fed to them, without prior knowledge of underlying physical laws. It is investigated whether a neural network architecture exits that can act as an alternative to the classical reservoir simulation to model the short-term operation of an underground gas storage facility. An accuracy of the wellhead pressure prediction in the order of ±1 Bar is the prerequisite to the optimization of operation schedules. Data from an actual underground gas storage plant, operated by the Rohöl-Aufsuchungs AG (RAG) is the basis for all neural network simulations. Several neural network types are tested with regards to their suitability of producing adequate wellhead pressure predictions during injection and production of natural gas. A prerequisite towards the application of these networks is an appropriate data management workflow. This includes the treatment of missing samples and outliers in the measured data as well as the transformation of the irregularly sampled signal to a regular one. Two different resampling methods are built to ensure a clean signal is used to train the neural networks. All networks are trained using equal history matching and testing intervals coming from one data set. The results on the history matching and testing data set showed that there were some network architectures that could successfully model the wellhead pressures during gas storage operation. Good results were achieved when using recurrent neural network architectures, like the fully recurrent one, the Elman network and NARX model. The best model outputs on the history matching as well as the independent test interval were produced by the fully recurrent network. Although none of the models trained and tested could reach the desired accuracy of ±1 Bar, it could be shown that these recurrent neural networks were capable of modelling the wellhead pressures in general. Finally it can be said that neural networks could be a viable alternative if additional improvements towards an enhanced accuracy are made.

AB - The underground storage of natural gas and the security of supply thereof is an important issue for a country like Austria, which depends on imports of fossil fuels. Typically the operation mode of an underground gas storage facility is determined by the seasonal swing in the demand of gas opposed by the nearly constant supply through pipelines. A more short-term oriented operation schedule supported by a numerical simulation of various scenarios could improve the energy efficiency and reduce wear through optimized switching cycles. This master's thesis evaluates the application of neural networks to the simulation of gas storage behaviour. Neural networks are heuristic tools that can be used to model nearly any trend in the data fed to them, without prior knowledge of underlying physical laws. It is investigated whether a neural network architecture exits that can act as an alternative to the classical reservoir simulation to model the short-term operation of an underground gas storage facility. An accuracy of the wellhead pressure prediction in the order of ±1 Bar is the prerequisite to the optimization of operation schedules. Data from an actual underground gas storage plant, operated by the Rohöl-Aufsuchungs AG (RAG) is the basis for all neural network simulations. Several neural network types are tested with regards to their suitability of producing adequate wellhead pressure predictions during injection and production of natural gas. A prerequisite towards the application of these networks is an appropriate data management workflow. This includes the treatment of missing samples and outliers in the measured data as well as the transformation of the irregularly sampled signal to a regular one. Two different resampling methods are built to ensure a clean signal is used to train the neural networks. All networks are trained using equal history matching and testing intervals coming from one data set. The results on the history matching and testing data set showed that there were some network architectures that could successfully model the wellhead pressures during gas storage operation. Good results were achieved when using recurrent neural network architectures, like the fully recurrent one, the Elman network and NARX model. The best model outputs on the history matching as well as the independent test interval were produced by the fully recurrent network. Although none of the models trained and tested could reach the desired accuracy of ±1 Bar, it could be shown that these recurrent neural networks were capable of modelling the wellhead pressures in general. Finally it can be said that neural networks could be a viable alternative if additional improvements towards an enhanced accuracy are made.

KW - Reservoir Simulation

KW - Neural Networks

KW - Recurrent Networks

KW - Data Analysis

KW - Resampling

KW - Gaussian Filtering

KW - Underground Gas Storage

KW - Wellhead Pressure

KW - Flow Rate

KW - Forecast

KW - Numeric Modelling

KW - Lagerstättensimulation

KW - Neuronale Netze

KW - Datenanalyse

KW - Abtastratenkonvertierung

KW - Gaussfilter

KW - Untertageerdgasspeicher

KW - Sondenkopfdruck

KW - Vorhersage

KW - Modellrechnung

M3 - Master's Thesis

ER -