Fig. 3. Hardware demonstration for on-device load forecasting in our framework.

a Schematic of the hardware platform. The established platform instantiates the proposed end-edge-cloud framework for comprehensive experiments. The memory-constrained smart meter cannot match the requirement of numerous variable parameters and massive amounts of constantly collected data for model training. b Comparison of forecasting accuracy versus memory usage on smart meters for our method and benchmark methods with different model sizes. The average forecasting accuracy with 95% confidence intervals is presented with five independent experiments. Source data are provided as a Source Data file.