At present, finite element modeling or laboratory test methods are often used in the research of the overall stability of high-strength steel members. However, the prediction method based on machine learning has greatly improved the accuracy and convenience of component performance prediction. To accurately predict the overall stability of welded constant section box columns made of high strength steel, a method of constructing a database based on the fiber model and establishing prediction models by machine learning is proposed in this paper. Firstly, the input and output parameters of the model are determined, and the database is established by the fiber model method. Then, three different types of machine learning models and empirical models in the existing specifications are selected for prediction, and the performance is compared according to the evaluation index. Finally, the rationality of machine learning models is analyzed according to interpretable algorithms. The results show that the prediction results of most machine learning models are in good agreement with the experimental results, which are slightly higher than the empirical models in the existing codes, and the Gaussian process regression model has the best prediction performance for the overall stability of high-strength steel members; The influence trend of various parameters on the overall stability of components meets the expectation, which verifies the rationality and reliability of the machine learning model; The regularized slenderness ratio has the greatest influence on the prediction results, while the initial defects have the least influence.