Time series forecasting algorithm based on federated learning
To meet the constantly evolving demands for data privacy protection,a time series forecasting algorithm designed for distributed scenarios was proposed.The main improvements of the algorithm were as follows:during the local model training phase at client nodes,a regularization term was employed to constrain the training direction of local model,resolving the issue of local model drift;in the global model aggregation phase,a client contribution estimation strategy was proposed,which allocated weights based on the extent of client contributions to ensure fairness in client collaboration and enhance the generalizability of global model.To validate the effectiveness of improved algorithm,it was compared with the baseline federated learning algorithm FedAvg on ETTh1,ETTm1,and Weather datasets.Experimental results showed that the improved algorithm reduced the mean square error EMS by 2.99%on ETTh1 dataset and 3.57%on ETTm1 dataset.Incorporating the regularization term and client contribution estimation strategy into the algorithm led to respective decreases of 0.84%and 2.78%in EMS,and a combined decrease of 3.03%in EMS,confirming that the proposed algorithm showed higher predictive accuracy.
federated learningmachine learningtime series forecastingdistributed systemsdeep learning
NETL
NSTL
万方数据
