Introducing edge intelligence to smart meters via federated split learning

Yehui Li, Dalin Qin, H. Vincent Poor, Yi Wang

Abstract

The ubiquitous smart meters are expected to be a central feature of future smart grids because they enable the collection of massive amounts of fine-grained consumption data to support demand-side flexibility. However, current smart meters are not smart enough. They can only perform basic data collection and communication functions and cannot carry out on-device intelligent data analytics due to hardware constraints in terms of memory, computation, and communication capacity. Moreover, privacy concerns have hindered the utilization of data from distributed smart meters. Here, we present an end-edge-cloud federated split learning framework to enable collaborative model training on resource-constrained smart meters with the assistance of edge and cloud servers in a resource-efficient and privacy-enhancing manner. The proposed method is validated on a hardware platform to conduct building and household load forecasting on smart meters that only have 192 KB of static random-access memory (SRAM). We show that the proposed method can reduce the memory footprint by 95.5%, the training time by 94.8%, and the communication burden by 50% under the distributed learning framework and can achieve comparable or superior forecasting accuracy to that of conventional methods trained on high-capacity servers.

The Challenge: The "SRAM Wall" and Privacy

While modern smart meters are deployed globally, they are largely limited to acting as data loggers. These devices typically utilize low-power microcontrollers, such as the ARM Cortex-M4, which possess extremely limited Static Random-Access Memory (SRAM)—often as little as 192 KB. This hardware constraint creates a "memory wall" that prevents the local training of sophisticated deep learning models required for accurate load forecasting.

Beyond hardware, two other significant hurdles exist:

  1. Privacy: High-resolution energy data can reveal sensitive information about household routines, making consumers hesitant to share raw data with central servers.
  2. Network Congestion: Transmitting raw data or massive model updates from billions of devices to the cloud creates immense communication overhead and latency.

The Federated Split Learning (FSL) Framework

To address these challenges, the researchers developed an end-edge-cloud federated split learning (FSL) framework. This system synergizes two distributed learning paradigms to bypass hardware limits while maintaining data sovereignty.

The End-Edge-Cloud Federated Split Learning (FSL) framework architecture.

Technical Pillars

  • Efficiency-Optimal Model Splitting: The neural network is partitioned into segments. The smart meter (the "end") handles only initial feature extraction and final regression, while the resource-intensive "feature processor" resides on a more capable edge server.
  • Parallel Training with Knowledge Distillation: Standard split learning is often sequential and slow. By introducing an auxiliary regressor on the meter, the system allows the meter and edge server to train in parallel. Knowledge distillation is then employed to synchronize the meter's lightweight local model with the larger edge-side model.
  • Semi-Asynchronous Aggregation: To prevent slow devices ("stragglers") from delaying the network, meters are clustered by hardware performance. Updates are synchronized within clusters at the edge but updated asynchronously to the cloud.

Experimental Validation

The framework was tested using a custom hardware testbed simulating a real-world grid: a tower server (cloud), three PCs (edge), and thirty ARM Cortex-M4 microcontrollers (smart meters). Two major datasets were used for validation: the Building Data Genome 2 (BDG2) and the Commission for Energy Regulation (CBTs) residential load data.

Performance and Comparative Analysis

The FSL framework demonstrates that "edge intelligence" does not require a compromise in performance. When compared against traditional centralized, local, and standard federated learning methods, FSL provides a superior balance of efficiency and accuracy.

Radar chart comparing Federated Split Learning against traditional learning paradigms.

Key Results

  • Resource Efficiency: The method achieved a 95.5% reduction in memory footprint and a 94.8% reduction in training time.
  • Communication: Data transfer burdens were slashed by 50% compared to standard distributed frameworks.
  • Grid Impact: Simulations indicated that this improved local forecasting could lead to a 31.79% reduction in electricity costs and a 59.78% reduction in carbon emissions by optimizing the management of local solar energy and storage.

Conclusion and Future Outlook

This research proves that advanced AI can be brought to the absolute edge of the power grid, even on devices with severe memory constraints. By intellectualizing smart meters, utilities can enhance grid flexibility and sustainability without compromising consumer privacy. Future work will focus on handling model heterogeneity—adapting models to individual household patterns—and incorporating the physical topology of the power grid into the clustering algorithms.

Reference: Li, Y., Qin, D., Poor, H. V., & Wang, Y. (2024). Introducing edge intelligence to smart meters via federated split learning. Nature Communications, 15:9044. https://doi.org/10.1038/s41467-024-53352-9

Summary updated

Abstract

The ubiquitous smart meters are expected to be a central feature of future smart grids because they enable the collection of massive amounts of fine-grained consumption data to support demand-side flexibility. However, current smart meters are not smart enough. They can only perform basic data collection and communication functions and cannot carry out on-device intelligent data analytics due to hardware constraints in terms of memory, computation, and communication capacity. Moreover, privacy concerns have hindered the utilization of data from distributed smart meters. Here, we present an end-edge-cloud federated split learning framework to enable collaborative model training on resource-constrained smart meters with the assistance of edge and cloud servers in a resource-efficient and privacy-enhancing manner. The proposed method is validated on a hardware platform to conduct building and household load forecasting on smart meters that only have 192 KB of static random-access memory (SRAM). We show that the proposed method can reduce the memory footprint by 95.5%, the training time by 94.8%, and the communication burden by 50% under the distributed learning framework and can achieve comparable or superior forecasting accuracy to that of conventional methods trained on high-capacity servers.
BibTeX
@inproceedings{li2024introducing,
  author = {Yehui Li and Dalin Qin and H. Vincent Poor and Yi Wang},
  title = {{Introducing edge intelligence to smart meters via federated split learning}},
  booktitle = {Nature Communications},
  year = {2024},
  doi = {10.1038/s41467-024-53352-9},
}
Figure