A Spatio‐Temporal Dynamics Model for Indoor Environmental Forecasting

Accurate forecasting of indoor environmental conditions is essential for intelligent building operations and comfort‐oriented controls. However, reliable prediction remains challenging due to spatial heterogeneity across sensing locations, complex multi‐modal interactions, and nonstationary temporal dynamics. This study proposes the Spatio‐Temporal Dynamics Model for Indoor Environmental Forecasting (STD‐IEF). STD‐IEF represents each sensor as a coordinate‐conditioned token and integrates distance‐aware spatial attention, sliding‐window temporal attention, and a horizon‐conditioned decoding mechanism to efficiently capture spatio‐temporal dependencies. Experiments were conducted using a high‐resolution dataset collected from 22 distributed sensor units in an office‐like room (18 sensors arranged in a planned 3D grid for model training and 4 sensors reserved at held‐out locations for generalization evaluation), measuring temperature, relative humidity, and wind speed. Using the latest 15 min of observations (aggregated to 1‐min resolution), the model predicts the subsequent 15 min. STD‐IEF achieves mean absolute errors of 0.185 °C for temperature, 2.231 %RH for relative humidity, and 0.016 m/s for wind speed for the 18 fixed‐layout sensors, and 0.219 °C, 2.346 %RH, and 0.017 m/s for the four held‐out sensors at unseen locations. The proposed framework remains computationally lightweight, supporting real‐time inference and making it suitable for future smart‐building applications.