Application of the hybrid neural network model for energy consumption prediction of office buildings

Accurate building energy consumption prediction is crucial to the rational planning of building energy systems. The energy consumption of buildings is influenced by various elements and is characterized by non-linearity and non-stationarity. To fully tap the time series characteristics of building energy consumption and heighten the model's prediction accuracy, this paper proposes a hybrid neural network prediction model combining attention mechanism, Bidirectional Gate Recurrent Unit (BiGRU), Convolutional Neural Networks (CNN), and the residual connection. The model uses BiGRU to train the extracted feature vectors by CNN on a two-way cycle. The attention mechanism highlights the key information extracted, and the residual connection is used to learn the features fully. Taking the energy consumption data of an office building in Guangzhou, China, as the object of study, the results indicate that the proposed model shows a stronger prediction accuracy than the commonly used model with an R2 of 90.74% and a CV-RMSE of 19.24%. Compared with the other five common models, the RMSE, MAPE, and MAE of the proposed model achieve lower error rates. Besides, the length 24 of the sliding window exceeds other lengths in the established model. The prediction accuracy of the established model in working hours outperforms the non-working hours of the office building. Building energy consumption prediction in the same season is better than that in the whole year.