Abstract: The thermal performance of gas turbines under variable working conditions exhibits strong non-linear characteristics. Calculations based on physical models are time-consuming and have larger calculation errors, while data-driven methods depend heavily on the quantity of sample data and lack constraints from physical mechanisms during the calculation process. To overcome the shortcomings of both approaches, the constraints among the internal and external influencing parameters, power generation efficiency, exhaust temperature and exhaust flow rate under variable working conditions were extracted from the physical mechanism of the compression and expansion processes of small gas turbines. Combined with the basic framework of neural network, a neural network model for small gas turbines operated under variable working conditions was proposed based on mechanism modification. The results of specific calculations indicate that high-precision neural network model of gas turbines under variable working conditions can be established using known gas turbine component performance and some data samples. The relative error is less than 0.941% in the prediction of key parameters by the improved model. The average relative error of prediction is 0.583%, which is 70.122% of the prediction error of the purely data-driven neural network using the same samples, and 14.916% of the calculation error of the method based on the physical model.