Accurate building models are essential for developing efficient energy control systems, but creating these models is challenging. Data-driven approaches tackle the time consuming and complex calibration of models by learning building dynamics directly from sensor data. However, the lack of data from newly built buildings hinders the adoption of such methods. To address this data scarcity, transfer learning approaches can adapt a model trained on a source building to a target one with minimal data. This study evaluates different transfer learning strategies-full fine-tuning, partial fine-tuning, and ensemble methods-across three types of building thermal models. Experiments were conducted using data from the UMAR Unit of the NEST building, Dübendorf, Switzerland and the CityLearn environment using data from multiple U.S. cities. Results show that transfer learning enables models to achieve performance close to the Oracle model-trained directly on the target building's data-while significantly reducing data requirements. Notably, partial fine-tuning maintains similar accuracy at a lower computational cost while the ensemble method, which averages fine-tuned models from different sources, can even outperform the Oracle model.