A novel heat load prediction algorithm based on DLinear-PatchTST model

Heat load prediction, as a key part of smart district heating system, is crucial for optimizing energy management, improving system efficiency and reducing operation cost. However, the existing heat load prediction models cannot effectively capture the local semantic information of short-term fluctuations in the heat load sequence and the global information of long-term trends in the heat load sequence. This makes the heat load prediction accuracy difficult to meet the optimization and regulation demands. In this paper, a heat load prediction model based on Dlinear-PatchTST is proposed. The combination of seasonal trend decomposition in Dlinear and patchTST effectively captures the local fluctuations and long-term trends in the heat load sequence and improves the prediction accuracy. In this paper, a detailed comparison and analysis of the proposed model with SOTA models such as Transformer, RNN, CNN-LSTM, and Seq2seq-Attention is carried out using three actual operating datasets of heat exchange stations. The results show that the heat load prediction model based on Dlinear-PatchTST proposed in this paper exhibits higher prediction accuracy and generalization ability, and the mean absolute percentage error (MAPE) of the three groups of experimental predictions is less than 3.17%.