Abstract: The terrain elevation in high altitude areas with large elevation differences is complex and variable. Although traditional elevation fitting methods,such as linear functions,surface fitting,BP neural network and genetic algorithm improved neural network can achieve elevation fitting,their fitting accuracy in high-altitude areas with large elevation differences remains to be improved. To effectively fit the terrain elevation in high altitude areas with large elevation differences,this paper proposes an elevation fitting method based on deep learning,using the second order leveling measurement data of a railway control network in the western region. The method employs a multi-layer perceptron as the core model,and selects the suitable combination of optimizers and activation functions according to their characteristics,to capture the terrain features and elevation change patterns of the region,and achieve high-precision elevation fitting. The paper also analyzes the impact of different combinations of optimizers and activation functions on the model performance. The results show that the deep learning model outperforms the BP neural network and genetic algorithm improved neural network methods in elevation fitting in high-altitude areas with large elevation differences,with the lowest MSE,the smallest MAE,and the R² closest to 1. Among them,the combination of RAdam optimizer and GELU activation function performs the best. The elevation fitting method of the deep learning model has higher accuracy and better generalization ability,and can effectively adapt to the complex and variable terrain features of high-altitude areas with large elevation differences.

Key words: deep learning, elevation fitting, high altitude areas with large elevation differences, optimizer, activation function