Abstract: There is an order of magnitude difference in the occurrence of different classes of flares. This makes it difficult for conventional convolutional neural network-based flare prediction models to capture M,X class flare features, which leads to the problem of low accuracy of high level flare prediction. In this paper, we discuss different deep long-tail learning methods to improve the accuracy of flare forecasting by controlling the variables for the long-tail distribution phenomenon in flare forecasting. We try to improve the forecast performance of the model for M,X flares from the perspectives of training set optimization, loss function optimization, and network weight optimization. The experiments on SDO/HMI solar magnetogram data show that the accuracy of M,X class flare prediction is significantly improved by 53.10% and 38.50%, respectively, and the recall is increased by 64% and 52% compared with the baseline model trained by conventional methods. It shows that the treatment of the long-tailed distribution of data is crucial in the flare forecasting problem, and verifies the effectiveness of the deep long-tailed learning method. This method of improving the accuracy of tail class forecasts can be applied not only to the field of flare forecasting, but also can be transferred to the analysis of forecasting other typical events of space weather with long-tailed distribution phenomenon.