<p>Previous studies have shown that there are significant errors in temperature forecasts in convection-permitting numerical weather prediction models, causing biases in wind and precipitation forecasts. In this study, the main sources of temperature forecast errors that stem from the uncertainty of physical parameterization schemes were investigated. The temperature forecasts were evaluated based on daily predictions over a month-long period. Overall, for all tested schemes, the largest biases originated from errors in cloud forecasting. Key factors affecting both temperature and cloud fraction predictions within each physical scheme were examined. These included key radiation processes, shallow cumulus convection, land-surface types, and the utilization of an urban model. For the radiation scheme, the number of quadrature points affected the amount of incoming and outgoing radiation and was a crucial setting that influenced the temperature forecasts. The subgrid cloud parameterization and the order of streams in the radiation scheme further enhanced the forecast differences. For the shallow convection scheme, different parameter settings mainly altered the cloud fraction simulation in cloudy areas rather than in clear skies. The subsequent temperature forecast was sensitive to the cloud fraction simulation. Parameter tuning was needed prior to applying the shallow convection scheme. For different land-surface datasets and urban model applications, accurate identification of land-surface types is crucial as it reflects the real surface albedo. This accuracy significantly influences the reliability of temperature forecasts. In addition, the more realistic heat capacity settings of buildings in the urban model improved the temperature forecast at the city scale.</p>