Fig. 2. Overview of the end-edge-cloud framework for on-device load forecasting.

We develop a federated split learning approach under this framework, which mainly incorporates three phases: (1) model splitting, in which the cloud server splits the large model and assigns a small portion to smart meters and a larger portion to the edge servers; (2) model training, in which multiple smart meters collaborate with edge servers to train the complete model; (3) model aggregation, in which the trained models are hierarchically aggregated by the edge servers and the cloud server to update the global model.