A multimodal fusion method for ultra-short-term prediction of cooling loads in building energy systems
The current research on the ultra-short-term prediction of building energy system cold load is confined to constructing the input feature set of a single modality,which somehow undermines the accuracy of cold load prediction.To address this,we propose a multimodal fusion method for ultra-short-term prediction of building energy system cold load.First,to solve the problem of a single input feature set,based on the historical data of the total cooling load of the building energy system and the cooling load of each user unit,three input feature sets of sequence-like,image-like,and video-like modalities are built respectively.Then,according to the data structure characteristics of the three modal input features,three deep learning prediction models are built in a targeted manner,namely,a two-way gated loop unit,spatiotemporal neural network,and three-dimensional convolutional neural network,and the preliminary total cooling load prediction results under the three modal inputs are obtained.Finally,a multimodal fusion method based on stacking integrated learning is proposed to carry out secondary learning on the preliminary prediction results of each prediction model under the three modal inputs to obtain the final total cooling load prediction results.Our simulation results from a test with the actual load data of the energy system of Arizona State University show our method effectively improves the ultra-short-term prediction accuracy of cold load.
multimodal fusionstacking integrated learningcooling loadultra-short-term predictionbuilding energy systems
国家科技期刊平台
NSTL
万方数据
