A novel robust heating load prediction algorithm based on hybrid residual network and temporal fusion transformer model

The district heating system involves a wide range and large scale, and the change of heat load exhibits significant non-linear, multi-coupling characteristics due to the influence of meteorological, building thermal inertia of numerous system factors. Additionally, the problem of missing values in the actual heating system also causes the existing heat load prediction models to frequently have difficulties in extracting the complex trends of the historical heat load data and result in the lack of prediction accuracy, which is detrimental to the supply-demand balance and leads to severe environmental problems such as energy wastage. A hybrid heat load prediction model based on Residual Network and Temporal Fusion Transformer algorithms was proposed in this paper to address the complexity of heat load changes in district heating system. The excellent capability of Residual Network in mining historical information contributes to the effective imputation of missing values and the establishment of a sound data foundation for the subsequent construction of data-driven heat load prediction models. The Gated Residual Network, multi-head attention mechanism, and other functional units in the Temporal fusion transformer model were sufficient to adequately extract the features and patterns of the historical heat load data and achieve reliable heat load prediction. Detailed comparative experiments with the actual data from four heat exchange stations at the district heating system in Anyang were conducted in this paper to validate the performance of the proposed hybrid heat load prediction model based on Residual Network and temporal fusion transformer. The mean absolute percentage error of the heat load prediction results in the four sites were all below 0.05 in the comparison experiments. The error metrics of Mean absolute error, Mean square error, and Root mean square error were also lower than other state-of-the-art algorithms. The feasibility and effectiveness of this model in practical applications are verified through experiments, which provide a new research perspective and practical reference for the heat load prediction of district heating system.