Search

An accurate characterization and prediction of heat loads in buildings connected to a District Heating (DH) network is crucial for the effective operation of these systems. The high variability of the heat production process of DH networks with low supply temperatures and derived from the incorporation of different heat sources increases the need for heat demand prediction models. This paper presents a novel data-driven model for the characterization and prediction of heating demand in buildings connected to a DH network. This model is built on the so-called Q-algorithm and fed with real data from 42 smart energy meters located in 42 buildings connected to the DH in Tartu (Estonia). These meters deliver heat consumption data with a 1-h frequency. Heat load profiles are analysed, and a model based on supervised clustering methods in combination with multiple variable regression is proposed. The model makes use of four climatic variables, including outdoor ambient temperature, global solar radiation and wind speed and direction, combined with time factors and data from smart meters. The model is designed for deployment over large sets of the building stock, and thus aims to forecast heat load regardless of the construction characteristics or final use of the building. The low computational cost required by this algorithm enables its integration into machines with no special requirements due to the equations governing the model. The data-driven model is evaluated both statistically and from an engineering or energetic point of view. R2 values from 0.70 to 0.99 are obtained for daily data resolution and R2 values up to 0.95 for hourly data resolution. Hourly results are very promising for more than 90% of the buildings under study. • Data driven model for heat-load prediction of buildings is developed. • Real data from 42 buildings connected to the district-heating in Tartu is used. • Different building typologies with different energy demand profiles are used. • The present model is applied to hourly and daily data. • R2 for hourly data range between 0.47 and 0.95 and for daily data range between 0.70 and 0.99 [ABSTRACT FROM AUTHOR]