The deep neural networks (DNNs) have been widely adopted in non-intrusive load monitoring (NILM) to analyze appliance-level power usage based on aggregate power consumption. Recently, transfer learning has gained attention in NILM as it improves the generalization of DNN models, particularly when transferring knowledge from one real-world dataset to another. However, due to the significant cost and time required for real-world data collection, synthetic datasets have emerged as a viable alternative for NILM research. Nevertheless, it remains unclear whether a well-trained DNN model on synthetic data can be effectively transferred to a real dataset, and which types of networks perform well when using transfer learning strategies. In this paper, we explore these crucial issues by investigating the generalization of different DNNs between synthetic and real datasets using two transfer learning schemes: direct transfer learning and transfer learning with strategy. Our results demonstrate that the sequence-to-point network shows promise for the transfer strategy from a synthetic dataset to a real dataset, and its performance relies on the similarity of probability mass functions between the two datasets.