Building Energy Consumption Prediction Driven by Multi-Source Data and Performance Comparison of Hybrid Models

Building energy consumption prediction plays a critical role in improving energy efficiency and reducing operational costs. Effectively integrating multi-source heterogeneous data and leveraging deep learning techniques to develop accurate and reliable prediction models is of significant importance. This study employed a two-stage feature selection method based on the maximum information coefficient and Granger causality test to optimize the feature set for building energy consumption prediction. Subsequently， four hybrid deep learning models were constructed by combining temporal convolutional network （TCN）， gated recurrent unit （GRU）， and attention mechanism （Attention） in different architectures： TCN-GRU， TCN-Attention， GRU-Attention， and TCN-GRU-Attention. The performance of these models was then comparatively analyzed from both architectural and data-scale perspectives using multi-source dynamic datasets from five categories of buildings. The results demonstrated that the TCN-GRU hybrid model achieved the best overall performance， with prediction errors lower than all baseline models， validating the complementary strengths of TCN in capturing local temporal features and GRU in learning long-term dependencies. Additionally， experiments revealed that， compared to hybrid models， the TCN/GRU models exhibited more robust performance improvements as data volume increased.