Supercritical carbon dioxide critical flow model based on deep learning.

The break accident process of a supercritical carbon dioxide (SCO2) reactor system presents a transcritical phenomenon. Operating under high pressure and a wide parameter range, the SCO2 system introduces multiphase characteristics to the critical flow of carbon dioxide (CO2) at various system positions. Nevertheless, current research lacks a comprehensive critical flow model capable of accommodating a broad parameter range with high precision. Data-driven approaches offer the potential to enhance accuracy by leveraging an expanding training database. To precisely forecast SCO2 critical flow efficiency, this study established an SCO2 critical flow model using deep learning techniques. The conservation equations and sensitivity analysis of experimental data informed the feature selection for the deep learning model, accomplished through the recurrent neural network (RNN) method, employing K-fold cross-validation and L2 regularization. The result was the SCO2-RNN critical flow model, boasting improved prediction accuracy and enhanced generalization capabilities. Subsequently, the optimal hyperparameters were determined by using a genetic algorithm. This model yielded an average error of 4.88% in predicted results, with a maximum error of 14.24%. The average error when extrapolating to the generalization data was 5.73%, with a maximum error of 20.45%. After the implementation of transfer learning, the average error decreased to 1.75%, with a final maximum error of 4.15%. Generalization results, using new data based on the trained model, underscore that the deep learning model meets engineering requirements for both efficiency and accuracy. • A wide-parameter, high-precision critical flow model was constructed based on the RNN deep learning model. • K-fold cross-validation was used to overcome the problem of insufficient training data, whereas L2 regularization was used to prevent overfitting. • A genetic algorithm was used to optimize the hyperparameters, and transfer learning was introduced to expand the applicability of the model. • The average model error is as low as 1.68% and the maximum error is as low as 4.15%. [ABSTRACT FROM AUTHOR]