Urban Building Energy Modeling aims at assessing the building energy performance at city scale with as little computational effort as possible. Thus, different methods have been developed in the last years to reduce the required calculation time by simplifying the modeling approach, selecting only representative buildings, or minimizing the building description. Starting from the latter ones, this work proposes a novel algorithm capable of abstracting a randomly shaped building into a representative shoebox. The presented shoebox generation algorithm is based on a preliminary sensitivity screening analysis on a set of reference parallelepiped-shaped thermal zones. This allowed the identification of the most significant geometry indicators influencing the building’s performance. Based on this, more complex geometries have been simplified to the shoebox with the same indicators and the accuracy of the algorithm has been evaluated comparing the simulated performance of simplified and original buildings. The approach includes the definition of equivalent shading surfaces, to account for self-shading elements in the original building geometry. The algorithm has shown good accuracy not only on the hourly thermal loads, but also the zones’ hourly temperature profiles, reducing to one third the energy simulation time with respect to the detailed building model. Although not as fast as other urban modelling approaches in the literature, it can retain accurate results at a finer time scale, i.e., on hourly basis, which is necessary in applications such as district heating and energy networks.