Fast prediction model for indoor temperature distribution based on adaptive spatio-temporal graph convolutional network with limited sensors

The implementation of non-uniform indoor thermal environments serves as an effective approach to simultaneously enhancing occupant thermal comfort and building energy efficiency. The acquisition, prediction, and precise control of indoor environmental parameters are pivotal for addressing the diverse comfort requirements of occupants.

To enable real-time prediction of indoor temperature distribution, this study introduces an Adaptive Spatiotemporal Graph Convolutional Network (ASTGCN) model, which operates efficiently with limited sensors. A case study was conducted in an office room in Beijing, where data were collected from 39 measurement points at various spatial locations over a 10-day period. Experimental results indicate that the proposed model, during its operational phase, can accurately reconstruct the temperature distribution across the remaining 32 locations within the space using only 7 fixed sensors.

The model achieves a MAE of 0.34°C and completes computations within 8 seconds, demonstrating both high accuracy and computational efficiency.