A Seq2Seq Framework for Spatiotemporal Forecasting in Ventilation Systems

Rapid acquisition of flow field data is of significant importance in both scientific and engineering contexts. However, conventional computational fluid dynamics (CFD) methods require substantial computational resources and time to obtain high-resolution flow field information, particularly in transient scenarios. Deep neural networks (DNNs), by leveraging their strengths in solving nonlinear problems and bypassing the numerical iteration of governing physical equations, offer distinct advantages for large-scale flow field data acquisition. This paper proposes a sequence-to-sequence model for transient flow field prediction, incorporating an attention mechanism, to forecast future flow field states. The model employs a graph-based structure to accommodate unstructured data. It first encodes a known input sequence to extract latent representations, which are then decoded along with the flow field state from the previous time step to predict subsequent sequences. Results demonstrate that in two- dimensional cases, the trained model achieves a prediction error of approximately 3% for future flow sequences while significantly accelerating inference speed. This approach enables rapid estimation of dynamically changing flow fields in indoor environments, providing valuable references for ventilation research.