Due to the inherent complexity and irregularity of cold load time series data, problems such as gradient disappearance, modal aliasing and over-fitting are prone to occur during the prediction process. Predicting the cold load of large public buildings remains a challenging task. To solve this problem and improve the prediction accuracy, the VMD-GRU model is proposed in this study. Real data from large public buildings were utilized to test the proposed model. The prediction process involves the following steps: 1) Correlation analysis of the original data and selection of highly correlated predictors; 2) Decomposition of the original data sequence into independent eigenmode functions using VMD; 3) Prediction of each component using GRU ; 4) Aggregation of component prediction results to obtain the cold load prediction value. To validate the model's effectiveness, a large public building in Xi'an is taken as an example for energy consumption analysis. The results are compared with other prediction models, including BP, GRU, EMD-BP, VMD-BP, EMD-GRU. Experimental results show that the proposed model effectively solves the problems, such as gradient disappearance, modal aliasing and over-fitting, accurately predicting the cold load of large public buildings.