The invention discloses a federated learning training method based on model dispersion. The invention relates to the field of artificial intelligence in edge calculation. According to the invention, in a real environment, data are often non-uniform and are distributed in a non-independent same manner, and unbalanced distribution of the data enables model updating uploaded to a central server by each client to have different degrees of difference, so that a high-quality model is difficult to train by randomly selecting the clients to participate in training. Meanwhile, the unbalanced distribution of the data can also amplify the influence caused by over-fitting, and model divergence is caused when the influence is serious. According to the method, in order to train a high-quality model under the condition of data imbalance, an updating strategy of a dynamic loss function is adopted to improve the stability of the model, and a client is selected according to the importance of the model,so that the accuracy and convergence rate of the model are improved. Meanwhile, on the basis of the two, a large number of traversal times and a proper regularization parameter mu are selected, so that the performance of the model is optimal.