<p>Dissolved gas analysis in oil is widely regarded as an important diagnostic method for identifying potential faults inside power transformers. Carrying out prediction of gas production in oil helps to achieve early fault diagnosis of equipment. However, many intelligent models for gas concentration prediction are purely data-driven, so their predictive performance is often constrained by the scale and quality of the available data, and their generalization capability and interpretability are typically limited. To improve this problem, a prediction method incorporating prior physical knowledge is proposed for estimating gas generation concentration in transformer oil in this paper. Firstly, this paper builds a discharge gas production experimental platform to obtain the gas production law under the oil-paper insulation system and establishes the mathematical relationship between gas production and energy release. Then, based on the above physical mechanism, the constraint equation is constructed and embedded in the time-series model to achieve high-precision prediction of characteristic gases. Finally, the method described in this paper is compared and analyzed with other methods that are not embedded in physical laws. The results show that compared with traditional machine learning methods, the Physics-Informed Neural Network (PINN) model is not only superior in prediction performance, but also shows good generalization ability and interpretability under insufficient sample conditions.</p>