This paper proposes a novel deep learning framework called DL-GCN for spatial-temporal indoor temperature prediction in multi-zone buildings. The framework combines distributed Long Short-Term Memory (LSTM) networks with Graph Convolutional Networks (GCN) to capture spatial-temporal correlations and multivariable coupling features. The GCN is used to extract spatial features from temperature and humidity data, while the distributed LSTM module fuses other data such as light and AC power consumption to learn coupling interactions and temporal characteristics. The model is validated using real-world datasets from a large-scale building, demonstrating superior performance in multi-zone indoor temperature prediction. The key contributions include spatial information extraction with GCN, multivariable coupling feature extraction, and a new model architecture. The DL-GCN model outperforms existing methods in terms of Mean Absolute Error (MAE), Mean Square Error (MSE), and correlation coefficient. Ablation experiments show the effectiveness of spatial, temporal, and distributed information fusion modules. The model is effective in capturing spatial, temporal, and multivariable coupling features, leading to improved prediction accuracy for multi-zone indoor temperature in large-scale buildings. The study also highlights the importance of considering multivariable coupling in temperature prediction for optimal HVAC system regulation.This paper proposes a novel deep learning framework called DL-GCN for spatial-temporal indoor temperature prediction in multi-zone buildings. The framework combines distributed Long Short-Term Memory (LSTM) networks with Graph Convolutional Networks (GCN) to capture spatial-temporal correlations and multivariable coupling features. The GCN is used to extract spatial features from temperature and humidity data, while the distributed LSTM module fuses other data such as light and AC power consumption to learn coupling interactions and temporal characteristics. The model is validated using real-world datasets from a large-scale building, demonstrating superior performance in multi-zone indoor temperature prediction. The key contributions include spatial information extraction with GCN, multivariable coupling feature extraction, and a new model architecture. The DL-GCN model outperforms existing methods in terms of Mean Absolute Error (MAE), Mean Square Error (MSE), and correlation coefficient. Ablation experiments show the effectiveness of spatial, temporal, and distributed information fusion modules. The model is effective in capturing spatial, temporal, and multivariable coupling features, leading to improved prediction accuracy for multi-zone indoor temperature in large-scale buildings. The study also highlights the importance of considering multivariable coupling in temperature prediction for optimal HVAC system regulation.