Cakformer: Transformer model for long-term heat load forecasting based on Cauto-correlation and KAN

Accurate prediction of heat loads is crucial for ensuring stable operation of thermal systems and efficient planning of thermal resources. In recent years, different Transformer architectures have been used for heat load prediction. However, these models often result in unsatisfactory prediction accuracy and have poor interpretability in practical applications. To overcome these challenges, First, we propose a convolutional autocorrelation method to replace the traditional self-attention mechanism and increase the extraction ability of local information while extracting global time dependence. Then, Kolmogorov-Arnold network (KAN) was introduced instead of Multi-Layer Perceptron (MLP) to more accurately capture and represent complex relationships in high-dimensional data. In addition, Synchrosqueezed Wavelet Transforms (SWT) is used to reduce the noise of the data. To verify the performance of Cakformer., two datasets were used to predict future heat loads at 96, 192, 336, and 720 steps. The experimental results show that Cakformer has the best prediction performance and strong universality compared with the seven comparison models.