Combining heating degree days method and machine learning techniques: implementation of a new model for mapping buildings heating demand at urban scale

• Comparative study of two machine-learning models for heating demand quantification. • Both models rely on easily obtainable and limited parameters. • Building’s base temperature as an output of one of the models. • Simplicity and scalability enable urban and large-scale energy modelling. • Real case study demonstrates practical applicability of the developed heating models. Building’s energy renovation holds significant potential for reducing energy consumption in the building sector. Consequently, there is a growing momentum to extend the concept of nearly zero-energy buildings to entire districts, fostering the development of Net-Zero Energy Districts. This shows the need for large-scale urban energy modelling to identify key renovation areas and to provide an initial evaluation of the impact of different measures to be implemented, thus facilitating effective energy transition planning. This paper proposes the development of a simple model to obtain early-stage estimations of energy demand by determining a variable base temperature specific to each building’s characteristics, which can then be used to estimate energy demand using the degree-day method. The starting hypothesis is that a base temperature can be determined for each building based on its characteristics (thus becoming an indicator of its energy performance) and that this temperature can be determined from publicly available input variables. To this end, a black box model was developed to calculate heating and cooling demand in buildings, combining the Heating Degree Days (HDD) concept with machine learning techniques to obtain a variable base temperature from a limited number of building parameters. Thus, different parameters affecting heating and cooling demand, including geometric and thermal features as well as climate-related factors, were evaluated in detail, and the most significant were selected. A dataset of 50,000 virtual cases, generated by varying these parameters, was used to train and validate various models, with an additional 1,000 scenarios for testing. This study aims to describe the process followed to determine the models, compare them, select the best-performing one and demonstrate its district-scale implementation using a real case study in northern Spain. Substantial input data were obtained from public databases. The results shown a strong predictive capacity, highlighting the model’s potential as an effective tool to support decision-making on energy renovation solutions. By providing accurate initial heating demand estimations, it overcomes the limitations of traditional HDD-based approaches while maintaining simplicity and ease of use