According to the U.S. Energy Information Administration's 2021 Annual Energy Report, consumption of many energy sources has declined by the COVID-19 in 2020. However, despite this energy trend, natural gas has recorded a rapid increase, unlike oil, and its consumption is expected to continuously inc...
According to the U.S. Energy Information Administration's 2021 Annual Energy Report, consumption of many energy sources has declined by the COVID-19 in 2020. However, despite this energy trend, natural gas has recorded a rapid increase, unlike oil, and its consumption is expected to continuously increase until 2050. In addition, conforming to the Paris Agreement in 2015, countries are making great efforts to reduce greenhouse gas emissions. As one of the efforts, hydrogen has been mentioned as an eco-friendly energy source without greenhouse gas emissions. In the opinion of the International Energy Agency, in 2021, global hydrogen demand is expected to grow steadily and show a seven-times increase compared with the current value by 2070. Therefore, natural gas and hydrogen might become key energy sources in the future global energy market. Most of the gas is transported through pipelines, which is known as the most cost-effective method for long-distance transport. However, material defects, corrosion, abrasion, and environmental influences can cause leak problems and can incur serious human, material, and environmental damage. In practice, a pipeline leak is one of the most common accidents in a gas pipeline, and various studies have been conducted to rapidly and accurately detect the leak. In addition, with the rapid growth of computing technology, research on safety monitoring and management using machine learning technology is being actively conducted. However, previous studies are insufficient in actual field applicability and leak detection in transient conditions.
Therefore, this study aimed to develop machine learning models that can detect leak size and location in steady and transient state conditions, respectively. The steady state leak detection model was made for a subsea natural gas pipeline operating in the Bay of Bengal and the transient model was developed for an onshore methane-hydrogen blending gas pipeline in Canada. The flow analysis was performed using Schlumberger's OLGA software, which is widely used in the industry and can analyze fluid flow for both steady and transient state conditions.
For the development of a steady state leak detection model, a base case was constructed using field data and matched with production history. By using the base model, various leak scenarios were generated and flow characteristics, sensitive variables, steady state arrival time, and flow pattern were analyzed. As a result, mass flow, pressure, and temperature were selected as the most sensitive parameters, and all cases arrived in a steady state within 10 minutes. In addition, a parametric study was performed by changing the leak size and location for obtaining machine learning training data. The mass flow rate, pressure, and temperature values were collected at the pipeline inlet and outlet points and they were used as machine learning input data. Multilayer perceptron (MLP), one of the deep learning techniques, was applied as a machine learning algorithm and an optimal learning model was derived through hyperparameter adjustment. The model accuracy was evaluated using and Mean Absolute Error (MAE). As result, the leak size detection showed very high accuracy, while the leak location showed relatively low accuracy. Therefore, data quality improvement and screening were carried out to improve the accuracy of the leak location, and the results were improved by 80% compared to the initial detection model.
Next, a long-distance onshore pipeline case was constructed for a methane-hydrogen blending fluid. As a result of the base case simulation, mass flow and pressure were most affected when the leak occurred, and it took 6 hours to arrive in a steady state. Therefore, it was necessary to analyze the flow characteristic in the transient state, and the trend data at the pipeline inlet and outlet with the leak size and location were analyzed. In particular, a wavelet transform method that can generate a two-dimensional image was applied to effectively express the trend data features over time. The converted image was used as training data for Convolution Neural Network (CNN), which is an image classification machine learning technique. The model accuracy was evaluated using classification accuracy and confusion matrix, and the leak location accuracy was relatively low similar to the steady state model. Therefore, the data acquisition time adjustment and data screening process were performed, and the accuracy was improved to be over 80%.
In conclusion, this study confirmed that rapid and accurate detection through machine learning is possible when a leak occurs in a steady and transient state condition. The result of this study is expected to be used complementarily with existing leak detection methods and it can help the operator`s correct and quick decision-making. In addition, this approach might be applied to most pipelines for fluid transport, and effective leak detection might also be possible.
According to the U.S. Energy Information Administration's 2021 Annual Energy Report, consumption of many energy sources has declined by the COVID-19 in 2020. However, despite this energy trend, natural gas has recorded a rapid increase, unlike oil, and its consumption is expected to continuously increase until 2050. In addition, conforming to the Paris Agreement in 2015, countries are making great efforts to reduce greenhouse gas emissions. As one of the efforts, hydrogen has been mentioned as an eco-friendly energy source without greenhouse gas emissions. In the opinion of the International Energy Agency, in 2021, global hydrogen demand is expected to grow steadily and show a seven-times increase compared with the current value by 2070. Therefore, natural gas and hydrogen might become key energy sources in the future global energy market. Most of the gas is transported through pipelines, which is known as the most cost-effective method for long-distance transport. However, material defects, corrosion, abrasion, and environmental influences can cause leak problems and can incur serious human, material, and environmental damage. In practice, a pipeline leak is one of the most common accidents in a gas pipeline, and various studies have been conducted to rapidly and accurately detect the leak. In addition, with the rapid growth of computing technology, research on safety monitoring and management using machine learning technology is being actively conducted. However, previous studies are insufficient in actual field applicability and leak detection in transient conditions.
Therefore, this study aimed to develop machine learning models that can detect leak size and location in steady and transient state conditions, respectively. The steady state leak detection model was made for a subsea natural gas pipeline operating in the Bay of Bengal and the transient model was developed for an onshore methane-hydrogen blending gas pipeline in Canada. The flow analysis was performed using Schlumberger's OLGA software, which is widely used in the industry and can analyze fluid flow for both steady and transient state conditions.
For the development of a steady state leak detection model, a base case was constructed using field data and matched with production history. By using the base model, various leak scenarios were generated and flow characteristics, sensitive variables, steady state arrival time, and flow pattern were analyzed. As a result, mass flow, pressure, and temperature were selected as the most sensitive parameters, and all cases arrived in a steady state within 10 minutes. In addition, a parametric study was performed by changing the leak size and location for obtaining machine learning training data. The mass flow rate, pressure, and temperature values were collected at the pipeline inlet and outlet points and they were used as machine learning input data. Multilayer perceptron (MLP), one of the deep learning techniques, was applied as a machine learning algorithm and an optimal learning model was derived through hyperparameter adjustment. The model accuracy was evaluated using and Mean Absolute Error (MAE). As result, the leak size detection showed very high accuracy, while the leak location showed relatively low accuracy. Therefore, data quality improvement and screening were carried out to improve the accuracy of the leak location, and the results were improved by 80% compared to the initial detection model.
Next, a long-distance onshore pipeline case was constructed for a methane-hydrogen blending fluid. As a result of the base case simulation, mass flow and pressure were most affected when the leak occurred, and it took 6 hours to arrive in a steady state. Therefore, it was necessary to analyze the flow characteristic in the transient state, and the trend data at the pipeline inlet and outlet with the leak size and location were analyzed. In particular, a wavelet transform method that can generate a two-dimensional image was applied to effectively express the trend data features over time. The converted image was used as training data for Convolution Neural Network (CNN), which is an image classification machine learning technique. The model accuracy was evaluated using classification accuracy and confusion matrix, and the leak location accuracy was relatively low similar to the steady state model. Therefore, the data acquisition time adjustment and data screening process were performed, and the accuracy was improved to be over 80%.
In conclusion, this study confirmed that rapid and accurate detection through machine learning is possible when a leak occurs in a steady and transient state condition. The result of this study is expected to be used complementarily with existing leak detection methods and it can help the operator`s correct and quick decision-making. In addition, this approach might be applied to most pipelines for fluid transport, and effective leak detection might also be possible.
주제어
#Natural gas Hydrogen blending gas Pipeline flow simulation Multilayer perceptron MLP Convolution neural network CNN Leak detection 천연가스 수소 혼합가스 관 유동 시뮬레이션 다 층퍼셉트론 합성곱 신경망 파이프라인 누출진단
※ AI-Helper는 부적절한 답변을 할 수 있습니다.