Spatio-temporal prediction of indoor thermal environment based on Graph Neural Networks and Recurrent Neural Networks

Indoor Environmental Quality (IEQ) significantly impacts occupant comfort and productivity, particularly in open-plan offices where spatial heterogeneity poses challenges for thermal environment regulation. Existing data-driven models primarily focus on temporal dependencies, overlooking spatial interactions critical for accurate thermal predictions.

This study proposes a novel Graph Neural Network-Long Short-Term Memory (GNN-LSTM) model that integrates spatial topology with temporal dynamics to enhance prediction accuracy. A case study in an open-plan office demonstrates that the model achieves high accuracy (R2 = 0.89), outperforming conventional Recurrent Neural Networks (RNN). Additionally, GNN-LSTM exhibits greater stability and generalization capability, ensuring robust performance across diverse spatial configurations.

The findings provide a foundation for intelligent thermal regulation in large-scale office spaces.