Long-term forecasting of hourly district heating loads in urban areas using hierarchical archetype modeling

This paper demonstrates and validates the application of a recently proposed archetype modeling and calibration framework for setting up 11 stochastic archetype building energy models of Danish detached single-family houses (SFH's). For this task, the municipal district heating system of Aarhus, Denmark, and its associated building stock were employed as case study, together comprising a dataset of 18 475 SFH's with hourly time series heating data for two years (2017-2018). The 11 physics-based archetype models were each calibrated using a training building sample with data from a one-year calibration period (2017) and tested for their ability to forecast the heat load of another building sample in a previously unseen one-year validation period (2018). The calibrated archetype models were further tested for their joint forecasting ability to match the aggregated heat load of six suburban dwelling areas of different composition and location within the city region of Aarhus, and finally for their ability to forecast the entire citywide dataset of 18 475 SFH's. The modeling framework performs very well for the aggregated citywide predictions with a practically non-existent bias of the overall heat load during the validation period (NMBE < 0.5%) and with only moderate inaccuracies present in hourly load predictions (MAPE < 12%). The high forecasting accuracy validates the application of the demonstrated archetype modeling framework for long-term urban-scale predictions; however, analysis of the time series errors indicate that the performance could be further improved by focusing on a better representation of the holiday periods and by ensuring the training data to be adequately informative to enable a good calibration of the model parameters. The simplicity of the archetype models coupled with the applied physics-based model structure makes the framework suitable for general energy planning purposes. By adapting a more dynamic model structure, it would also be possible to apply the framework for more complex analysis of, for instance, the urban-scale demand response and the general heating flexibility of the building stock. (C) 2020 Elsevier Ltd. All rights reserved.