Aiming at increasing the accuracy and reliability of irradiance prediction, the paper proposes a prediction model of short-term irradiance based on improved Stacking ensemble learning and error correction. Firstly, through gradient boosting decision tree, the paper characteristically screens the original data set, removes redundant characteristics, and increases prediction accuracy and computing efficiency; then, an improved Stacking irradiance prediction model is established. In accordance with difference in prediction accuracy of different prediction models in the primary layer under K-fold cross-validation, the paper weights the prediction results, and applies Box-Cox to transform and process the training set data input from the first layer to second layer of Stacking, so as to increase the normality and homoscedasticity of prediction. Finally, the paper extracts the historical prediction error data, and applies random forest to construct an error model to further increase the prediction accuracy. Comparing with the traditional model and the classic Stacking model, the experimental results show that there are significantly improvements on the prediction performance of model.