In the event of a hydrogen leak, the build-up of hydrogen near the ceiling of an underground garage poses a significant safety risk. Fast and accurate estimation of hydrogen concentration distribution is crucial for risk assessment. This study proposes a novel neural network named multi-expert variational hybrid network (MEVHN) to predict the distribution of hydrogen concentration under the ceiling when the peak concentration reaches its maximum value during a leakage event. The model utilizes data from discrete sensors to make predictions. It incorporates a mixture of experts (MoE) framework to transform the sensor data into latent variables, which are then used by a variational autoencoder (VAE) decoder to predict the hydrogen concentration distribution. Constraints are added to the loss function to improve the prediction accuracy further. The results show that the MEVHN has an inference time of 1.3 seconds, a coefficient of determination (R²) of 0.977, a mean absolute error (MAE) of 1.86E-3, and a mean squared error (MSE) of 3.15E-5. These results indicate that the model performs well in predicting the 2D hydrogen concentration distribution.
Read full abstract