Federal learning-based time series model uncertainty correction method and system

The invention provides a federated learning-based time sequence model uncertainty correction method and system, and the method comprises the steps: aggregating the local label data distribution of each local client through a prediction model of the local client, and obtaining a global model covering the whole data distribution; performing calculation based on the data set and the global model to obtain prior distribution, learning conditional distribution on a potential space by using the prior distribution and the global model to obtain a posterior probability, guiding prediction model training by using the posterior probability, and obtaining an uncertainty model by using a feature combination and trained prediction model; modeling is performed on distribution of a time sequence regression task and a classification task by using heterovariance generalized Gaussian distribution, an uncertainty model is updated, and training of the uncertainty model is completed in combination with a loss function. According