Effective prediction of shale gas production is crucial for strategic reservoir development. However, in new shale gas blocks, two main challenges are encountered: (1) the occurrence of negative transfer due to insufficient data, and (2) the limited interpretability of deep learning (DL) models. To tackle these problems, we propose a novel transfer learning methodology that utilizes domain adaptation and physical constraints. This methodology effectively employs historical data from the source domain to reduce negative transfer from the data distribution perspective, while also using physical constraints to build a robust and reliable prediction model that integrates various types of data. The methodology starts by dividing the production data from the source domain into multiple subdomains, thereby enhancing data diversity. It then uses Maximum Mean Discrepancy (MMD) and global average distance measures to decide on the feasibility of transfer. Through domain adaptation, we integrate all transferable knowledge, resulting in a more comprehensive target model. Lastly, by incorporating drilling, completion, and geological data as physical constraints, we develop a hybrid model. This model, a combination of a multi-layer perceptron (MLP) and a Transformer (Transformer-MLP), is designed to maximize interpretability. Experimental validation in China's southwestern region confirms the method's effectiveness.