Abstract:

To ensure the safety of hydrogen-based fuels during production, storage, transportation and usage, the LEL test data of H₂, NH₃ and CH₃OH under different initial temperatures and pressures were collected, and a machine learning database consisting of 498 samples was constructed. Based on this dataset, feature correlation analysis was conducted, and the data was split into training and testing sets in an 8∶2 ratio. A five-fold cross-validation strategy was employed to enhance the stability and reliability of neural network model training. Multivariate linear regression models and deep neural network models were established separately. By comparing different activation functions, optimization algorithms, and numbers of neurons, the hyperparameters and network architecture of the neural network were determined. Concurrently, combustion reaction kinetics calculations were performed using a one-dimensional laminar premixed flame model and a detailed reaction mechanism, and the three methods were systematically compared and validated for accuracy. The results indicate that the optimal structure of the neural network is 5-10-12-1. The prediction performance is the best when using Rectified Linear Unit(ReLU) activation function and the adaptive moment estimation algorithm as the optimizer. The coefficient of determination R² of the model is 0.983 6, the root mean square error (RMSE) is 0.248 6, the mean absolute error (MAE) is 0.201 2, and the maximum MAE compared with the experimental values is 0.25%. The combustion reaction kinetics method has the second-best accuracy, with the maximum MAE of the experimental values being 0.58%. The multiple linear regression model has the lowest accuracy and poor generalization ability. R² of the model is 0.911 9, the RMSE is 0.296 5, and the MAE is 0.225 1, and the maximum MAE of the experimental values reaches 1.89%.

Key words: machine learning, hydrogen-based fuels, lower explosive limit(LEL), prediction model, neural network, combustion reaction kinetics