Efficient room-level heat load prediction in buildings using spatiotemporal distribution characteristics

Room-level heat load prediction is essential for optimizing building energy efficiency and managing intelligent HVAC systems. This study proposes a prediction method integrating spatiotemporal distribution features within a “global prediction + feature decomposition” framework to overcome traditional data-driven limitations. A spatial allocation matrix, based on building thermodynamics, quantifies heat load allocation across floors and orientations, while a temporal allocation matrix, derived from intermittent heating experiments, captures inter-room heat transfer effects. A thermodynamic model built with DesignBuilder and a ResGRU neural network enables overall heat load prediction, with spatiotemporal matrix decomposition ensuring rapid room-level estimations. Empirical results show that this method requires only total inlet heat metering for multi-scenario predictions, reducing sensor deployment while achieving 7.02 % accuracy. This approach supports zonal HVAC control and demonstrates strong potential for energy-efficient building applications.