Enhancing decomposition-based hybrid models for forecasting multivariate and multi-source time series by federated transfer learning

Time series forecasting is a complex task that demands both accuracy and efficiency. Hybrid models have shown promising forecasting performance. These models integrate decomposition algorithms (e.g., the multivariate variational mode decomposition, MVMD) with individual models such as the classical long shortterm memory network (LSTM). However, these models encounter challenges such as high time consumption in the model training process and data leakage issues when dealing with multivariate time series from different data owners (i.e., multi-source). To address these challenges, a new framework for hybrid models aimed at multivariate and multi-source time series (MMTS) forecasting, referred to as MVMD-x-FTL (e.g., MVMD-LSTMFTL), has been proposed. In this framework, MVMD refers to the federated MVMD (Fed-MVMD) algorithm we proposed with the federated transfer learning (FTL) technique, while the x represents any selected individual model. The framework is evaluated on 8 real-world datasets from various domains such as network traffic, solar radiation, energy consumption, etc. Compared to the classical framework MVMD-x (i.e., MVMD-LSTM), the proposed MVMD-LSTM-FTL framework reduces the LSTM model fitting time by an average of 56.17% and mean absolute error (MAE), root mean square error (RMSE), normalized RMSE (NRMSE) for forecasting accuracy by up to 1.79%. And the privacy budget epsilon in Fed-MVMD has no significant impact on the final forecasting results. These experimental results demonstrate that our proposed framework effectively addresses the concerns of time consumption and privacy leakage for multivariate and multi-source time series forecasting tasks.