A physical model driven machine learning for predicting maximum leakage rate in supercritical CO₂ release

Supercritical CO₂ pipelines are suitable to transport the large-scale CO₂ involved in Carbon capture and storage projects. The leakage process of supercritical CO₂ pipelines is accompanied by complex phase changes. Therefore, it is difficult to predict the maximum leakage rate accurately at present. In view of the shortcomings of traditional physical model methods such as complex modeling, too many assumptions and time-consuming calculations, a way of predicting the maximum leakage rate of supercritical CO₂ pipelines by machine learning method was proposed. It used to simply convolutional neural networks (CNN) and support vector machine improved by particle swarm optimization (PSO-SVM) respectively to study the leakage feature data generated by the isentropic choked flow leakage model. The prediction accuracy and generalization ability of the trained machine learning model were tested. The results show that: First, the average error between experimental data and prediction results of physical model, PSO-SVM, CNN is 28.82%. Second, the prediction accuracy of the two machine learning models shows little difference, the training time of CNN is much shorter than that of PSO-SVM, but the generalization ability of PSOSVM is stronger than that of CNN. Therefore, SVM is suitable for accurate prediction of small sample data, while CNN is more suitable for learning and prediction of large sample data. This study provides a new efficient method for predicting the maximum leakage rate of supercritical CO₂ pipelines.