Introducing Edge Intelligence to Smart Meters via Federated Split Learning

This is the codebase for our paper "Introducing Edge Intelligence to Smart Meters via Federated Split Learning".

Overview

Low-cost smart meters are ubiquitous in smart grids. Enabling resource-constraint smart meters to perform deep learning is quite challenging. Our end-edge-cloud framework reveals a new path for edge intelligence on smart meters. It improves 95.6% memory footprint, 94.8% training time, and 50% communication overhead, achieving comparable or even superior forecasting accuracy to that of conventional methods trained on high-capacity servers. We provide code for established experimental and simulation platforms.

Dataset

The current version supports the following datasets and tasks are saved in dataset:

• BDG2 - Building load forecasting
• CBTs - Residential load forecasting

Experimental platform

You can build the experimental platform with three components: microcontrollers, personal computers and tower server. experimental_platform is loaded with the code deployed on them.

Note: The microcontrollers are coded in C language and the computers and tower server are coded in Python.

To use the provided code, you are supposed to:

• Load the dataset dataset/BDG2_dataset/.csv or dataset/CBTs_dataset/.csv into microcontrollers.
• Compile experimental_platform/smart_meter/USER/.uvprojx and download the code to the flash memory of microcontrollers.
• Run experimental_platform/edge_server/Edge_server on computers.
• Run experimental_platform/cloud_server/Cloud_server on server.

Note: Please ensure the microcontroller is configured with at least 192 KB of SRAM and 1 GB of FLASH.

Note: Please ensure the communication network is connected and stable before use.

Simulation platform

You can also build the simulation platform with the tower server. simulation_platform is loaded with the code deployed on it.

To use the provided code, please run simulation_platform/Test.py to obtain the results:

server.centralized_train(): Cen

server.local_train(): Local

server.fed_train(): FedAvg

server.fed_prox_train(): FedProx

server.split_train(): Split

server.sflv1_train(): SFLV1

server.sflv2_train(): SFLV2

server.distillation_train(): Proposed

Experimental results

Enable resource-constrained smart meters to execute resource-intensive intelligent algorithms!

• Performance comparison

• Accuracy versus memory footprint

• Impacts on energy management

Video

• Smart meters are performing 4-hour-ahead load forecasting. The red line is the truth value and the blue line is the prediction value.

Requirement

Experimental platform:

• Smart Meter: μVision 5.3+ and STM32F405 MCU
• Edge Server: Python 3.6+ and Inter i7 12700 4.9GHz CPU
• Cloud Server: Python 3.6+ and NVIDIA GeForce RTX 3080 Ti 12GB GPU

Simulation platform:

• Python 3.6+
• PyTorch 1.4.0+
• NVIDIA GeForce RTX 3080 Ti 12GB GPU