<p>The identification of thermal load features is a prominent research area that aims to infer potential external thermal loads on the structure in a non-destructive way. Currently, data-driven methods are effective tools in this area. However, these methods often exhibit limited generalization because their accuracy depends heavily on the quality and coverage of the training data. To enhance the generalization of data-driven methods, this paper integrates domain adaptation transfer learning with a ConvLSTM neural network, thereby presenting a robust transfer learning framework. First, key parameters sensitive to thermal load features were extracted to build the dataset based on simulation and experimental results. Then, transfer learning was performed by embedding the ConvLSTM neural network within a domain adaptation method. The results demonstrate that, compared with a solely ConvLSTM model without transfer learning, the proposed method achieves lower prediction errors on the experiment dataset (diameter relative error decreased from 28.34% to 9.72%, power relative error decreased from 21.09% to 3.5%). Compared with fine-tuning transfer learning, the proposed method has better generalization capabilities. These improvement proves that the proposed method enhances generalization performance on datasets with significant differences in data distribution.</p>