FVQVAEformer: A spatiotemporal prediction framework for hydrogen leakage concentration fields in hydrogen production workshops

Severe consequences may arise from hydrogen leakage in confined spaces, such as hydrogen production workshop (HPW). Accurately predicting the temporal evolution of hydrogen concentration dispersion following a leakage event is crucial for hydrogen production systems and personnel safety. This study presents FVQVAEformer, an efficient deep-learning-based spatiotemporal framework for predicting hydrogen concentration evolution near the ceiling following a release. First, a Fourier vector quantized variational autoencoder (FVQVAE) integrated with a Fourier Neural Operator (FNO) is developed to effectively capture frequency-domain features of high-dimensional turbulent concentration fields and achieve low-dimensional compression. Second, a stacked multilayer perceptron (MLP) module is employed to reconstruct the global low-dimensional initial concentration field using only concentration measurements from five sparse sensors together with operating-condition information. A decoder-only transformer is then adopted to perform continuous long-horizon rollout of the flow field. The results show that the proposed model can accurately predict concentration-field evolution within 150 s after leakage, achieving coefficient of determination (R2) values up to 0.99; despite error accumulation, the R2 remains above 0.92 at t = 150 s. The computational efficiency is improved by several thousand times compared with conventional CFD, and the prediction accuracy surpasses existing baseline models. Overall, the proposed framework satisfies the rapid-response requirements for hydrogen-leak accidents in HPW and provides a theoretical basis for the safe application and wider deployment of hydrogen energy.