Thermal load prediction using surrogate models in district heating systems

The optimization of surrogate modelling used in building thermal load management of District Heating Network (DHN) layouts for heating load prediction is crucial to reducing the contribution of building energy consumption globally. However, state-of-the-art surrogate models often struggle to capture the dynamic thermal mechanics of DHNs that occur in hourly intervals without sacrificing their low computational cost. Consequently, the model modification should be expanded beyond the present literature scope of conventional archetypes or multi-surrogate structures only. To this aim, this work proposes a novel surrogate model construction methodology, which emphasizes bolstering surrogate model performance through the combination of a model archetype, a multi-surrogate structure using rule-based data splitting and a criterion-based restriction of training data. Applying the modelling methodology on a real substation shows that the predictive performance of surrogate models depends on dataset restriction on single-surrogate cases, across all archetypes. Contrarily, in multi-surrogate cases, it depends on the model archetype used, as well as insufficient data quality and thermal mechanics manifesting in certain periods of the operation of the DHN substation. In general, this novel methodology can assist in future heat load prediction endeavours, by detecting surrogate modelling limitations posed in each DHN case and encouraging further improvements.