For small sample wind tunnel test data of a high-rise building with rectangular section and different aspect ratios, several typical machine learning methods, including decision trees, BP neural networks, and ensemble generalized regression networks, are used to determine the overall wind pressure shape coefficient of the building. Comparing the accuracy of predictive models obtained from different machine learning methods, for training sets with small sample data, traditional machine learning methods are prone to distortion in the prediction of different model data distribution patterns. Through boosting and bagging integration principles, a generalized regression network based on second-order integration is proposed. This model exhibits high accuracy and better robustness, enhancing the ability to uncover patterns in data distribution. Using this model, the variation of the maximum along-wind overall shape coefficient with the aspect ratios is predicted for the high-rise building with rectangular section. The prediction results show that maximum shape coefficient along wind increases to the maximum value at aspect ratio of 2 and then decreases with increase of aspect ratio. When the aspect ratio is greater than 4, maximum shape coefficient remained almost constant around 1.35.