Abstract: Due to the complexity of external factors such as weather conditions and wind turbine control strategies, future forecasting data often deviates from the training data distribution, leading to a significant decline in the accuracy of ultra-short-term wind power prediction models. To address this issue, an ultra-short-term wind power forecasting method based on invariant learning was proposed. The method learned the mapping relationship between invariant features and power by jointly optimizeing the environment inference module and the invariant feature learning module, achieving robust modeling. The results show that compared with two benchmark models, the proposed method reduces forecasting error of the normalized root mean square error and normalized mean absolute error by an average of 1.19-1.30 and 0.41-0.68 percentage points respectively.