An improved Transformer with DBB and KAN for smart building energy consumption prediction

Energy building energy consumption prediction plays a crucial role in energy management and sustainability. While deep learning approaches, particularly Recurrent Neural Networks (RNNs) like LSTM and GRU, have shown promise in this field, they face challenges in capturing the complex periodic patterns in building energy data. Traditional Transformer architectures, with their simple feedforward network structures and fixed activation functions, often struggle to fully capture the complex interactions among these influencing factors. This paper presents an improved Transformer model for building energy consumption prediction by incorporating three novel components: Diversified Branch Block (DBB), Fine-grained Temporal Extraction module (FTE), and Kolmogorov-Arnold Network (KAN). Experimental results on two real-world datasets demonstrate the model’s effectiveness. On the Belgium dataset, it achieves an MSE of 0.018, MAE of 0.057, and ${\mathrm {R}}^{2}$ of 0.974, significantly outperforming traditional approaches like LSTM, GRU, and standard Transformer. On the BGD2 dataset, the model exhibits robust performance across different building types, achieving MSEs of 0.016, 0.013, and 0.015 with corresponding $\mathrm{R}^{2}$ values of 0.944, 0.953, and 0.962 for educational, office, and residential buildings respectively. Ablation studies confirm the contribution of each component: DBB enhances feature extraction, FTE captures fine-grained temporal patterns, and KAN improves the modeling of nonlinear relationships. The integration of these components produces a synergistic effect, resulting in improved performance in building energy consumption prediction tasks.