Atmospheric drag is the largest non-gravitational perturbation experienced by low-orbit satellites, and the main source of error in calculating atmospheric drag stems from inaccuracies in the empirical models of thermospheric density. Currently, these empirical models generally exhibit errors exceeding 30%. To enhance the prediction accuracy of these models, a calibration method for thermospheric density empirical models based on Segment Recurrent Neural Network (SegRNN) is proposed. This method employs the segmentation and parallelism strategies of SegRNN for model training and inference, mitigating the issues of error accumulation and gradient instability that arise from excessive iterations in traditional RNNs. By analyzing the relationship between atmospheric density and external environmental parameters such as Ap, F
10.7, and F
10.7a, an improved neural network architecture named "SegRNN with Residual Block" is proposed. This architecture introduces external environmental parameters as dynamic covariates and employs Residual Block to encode these covariates, thereby extracting density-related information for the prediction period and further enhancing the prediction accuracy of SegRNN. Finally, the density data derived from the onboard accelerometer of the GRACE (Gravity Recovery and Climate Experiment) satellite is used to calibrate the NRLMSISE 2.0 model. The results indicate that the original error of the NRLMSIS 2.0 model is 31.3%. After calibration with SegRNN, the error was reduced to 8.0%. By introducing dynamic covariates, the model error was further reduced to 7.2%. Ultimately, the error of the final calibrated model decreased by 24.1%, demonstrating significant calibration effects.