Abstract: Accurate prediction of heating load not only effectively reduces energy consumption but also improves the efficiency of the heating system and user comfort. To enhance the accuracy of heating load prediction, this study combines the Variational Mode Decomposition (VMD) algorithm with an improved Informer model for heating load prediction. Firstly, the VMD algorithm is used to decompose the heating load data, reducing its non-stationarity. Secondly, relative positional encoding is introduced in the Informer model to better capture the dependencies in the sequential data and avoid information leakage, replacing the absolute positional encoding. Furthermore, dilated causal convolution is adopted instead of regular convolution to increase the receptive field and enhance the extraction of local information. Comparative experiments with mainstream prediction models (GRU, LSTM, Transformer, and Informer) are conducted on multiple datasets. The experimental results demonstrate that the RDC-Informer model achieves an R² evaluation metric of 98.3%, which is 11.6%, 6.3%, 4.7%, and 2.6% higher than the comparative models, respectively. Additionally, the effectiveness of the dilated causal convolution is verified by increasing the convolution kernel, confirming the applicability and accuracy of the RDC-Informer model and providing a reference for further improvement in real-time smart heating.

Key words: heat supply load forecast, Informer, dilated causal convolution, relative position coding, VMD