Model‐based studies on urban heat islands can be seriously affected by errors in near‐surface air temperature (T2), especially if errors differ between cities and their rural surroundings. Furthermore, errors in T2 strongly depend on selected parameterisation schemes, in particular on the planetary boundary layer (PBL) scheme and the urban canopy model (UCM). We developed the Central Europe Refined analysis (CER), a dataset generated by dynamically downscaling a global atmospheric reanalysis with the Weather Research and Forecasting (WRF) model for Central Europe (30 km), Germany (10 km), and the region of Berlin‐Brandenburg (2 km). CER data were analysed to study urban–rural and intra‐urban differences in T2 for Berlin as well as to test the sensitivity of T2 against two different PBL schemes, a mosaic approach, and three UCMs with different levels of complexity. Results were evaluated using data from 22 weather stations. All tested configurations simulated T2 with small deviations from observations. The PBL schemes predominantly control the deviation of T2. From the tested PBL schemes, the Bougeault–Lacarrére scheme performed better than the Mellor–Yamada–Janjić scheme. The application of different UCMs and the mosaic approach also influenced the deviations, but not as strongly as the PBL schemes. The performance of the UCMs regarding the representation of intra‐urban and urban–rural differences showed that differences were largest when using a complex multi‐layer UCM. Overall, the simplest model showed lowest deviations. We conclude that more research on UCMs is required because complex UCMs showed potentials but did not outperform the simple slab model.