A deep learning approach for energy management systems in smart buildings towards a low-carbon economy

Abstract Addressing the issue of cold load prediction in building energy systems, a multi-modal fusion deep learning approach is proposed. This method constructs input feature sets of three different modalities: sequence-like, image-like, and video-like, and employs bidirectional gated recurrent units, spatiotemporal neural networks, and 3D convolutional neural networks. Additionally, this paper introduces a multi-modal late fusion strategy based on stacking ensemble learning. Experimental results demonstrate that this method performs exceptionally well in cold load prediction tasks, achieving an MAPE of 5.45%, and R2 of 95.25, which is crucial for the practical implementation of low - carbon building energy management.