The weld pool analytical solution is an ideal method to visualize the thermal behavior of the welding process, to clarify the physical relationship between the welding process parameters and the shape of the weld pool and to quickly calculate the penetration. In previous work, the weld pool analytical model designed for time-varying welding speed situation was derived. The experimental results show that the maximum error of the calculated penetration is 18.91%. In order to improve the computational accuracy of the analytical model, this paper proposes a calibration method based on the in-situ reconstructed weld pool surface. Firstly, the RES-BiSeNet model based on convolutional neural network is built to extract the features of the weld pool surface. Then, the calibration parameters are obtained to calibrate the weld pool analytical model. The accuracy of the calibrated analytical model was verified through the welding experiments. The results show that the maximum calibrated penetration error is reduced to 9.7%. Finally, the influence of the heat source parameters on analytical model is discussed.