pith. sign in

arxiv: 2604.19800 · v1 · submitted 2026-04-10 · 💻 cs.LG · cs.AI· cs.SY· eess.SY

On-Meter Graph Machine Learning: A Case Study of PV Power Forecasting for Grid Edge Intelligence

Jian Huang , Zixiang Ming , Yongli Zhu , Linna Xu This is my paper

Pith reviewed 2026-05-10 16:48 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.SYeess.SY
keywords graph neural networksPV power forecastingsmart metersedge intelligenceONNX runtimemicrogridGCNGraphSAGE
0
0 comments X

The pith

Graph neural networks for photovoltaic power forecasting can be deployed and run on smart meter hardware in microgrids.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper explores the application of graph machine learning models to forecast photovoltaic power generation directly on edge intelligent meters within a microgrid setting. It details the process of training GCN and GraphSAGE models and adapting them for deployment using ONNX and ONNX Runtime, including the creation of a custom operator for the graph convolutional network. The study uses real data from a village microgrid to compare model performance on both personal computers and the smart meter itself, confirming that the models execute successfully on the constrained hardware.

Core claim

The central discovery is that two graph machine learning models, GCN and GraphSAGE, can be successfully trained on PC and then deployed to run on smart meter devices for PV power forecasting in a microgrid, with a customized ONNX operator enabling the GCN computations on the edge device.

What carries the argument

A customized ONNX operator for GCN that allows graph neural network inference on the resource-limited smart meter hardware.

If this is right

  • Local execution on meters reduces reliance on central servers for microgrid management.
  • Real-time PV forecasts become available at the grid edge without network latency.
  • Similar graph-based models could be adapted for other forecasting tasks on smart meters.
  • The approach demonstrates feasibility of edge intelligence for renewable energy integration.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Deploying such models could enhance privacy by keeping data local on the meter.
  • Scalability to larger networks of meters might require further optimization of the custom operators.
  • Integration with other sensor data could improve forecasting accuracy beyond the current graph structure.

Load-bearing premise

The smart meter hardware provides sufficient computational resources for the graph models to run efficiently without major losses in accuracy or speed.

What would settle it

Observing that the models on the smart meter produce forecasts with substantially higher error rates or runtimes compared to the PC versions would disprove the successful deployment claim.

Figures

Figures reproduced from arXiv: 2604.19800 by Jian Huang, Linna Xu, Yongli Zhu, Zixiang Ming.

Figure 1
Figure 1. Figure 1: Illustration of ONNX and ONNX Runtime E. Customized Operator Although the built-in operators provided by ONNX can handle most neural network models, some models, like GCN, contain operations that lack corresponding operator mappings. This prevents these models from being exported to the ONNX format. To solve this problem, ONNX provides a mechanism for creating customized operators, allowing developers to d… view at source ↗
Figure 5
Figure 5. Figure 5: The real smart meter used in this paper F. Hardware and Software Specifications of the Smart Meter The hardware platform for GNN deployment, depicted in [PITH_FULL_IMAGE:figures/full_fig_p003_5.png] view at source ↗
Figure 3
Figure 3. Figure 3: Model export integration 3) Backend kernel implementation: Implement the specific computational logic for the customized operator within the inference engine [11] (as shown in [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 6
Figure 6. Figure 6: Flowchart regarding how to register the customized operator in ONNX [PITH_FULL_IMAGE:figures/full_fig_p004_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: The ONNX’s computation graph for the exported GCN model, grouped into functional blocks: Data Input, 2-layer GCN (with MyGcnOp), [PITH_FULL_IMAGE:figures/full_fig_p005_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: The process for deploying the GraphSAGE model [PITH_FULL_IMAGE:figures/full_fig_p005_10.png] view at source ↗
Figure 9
Figure 9. Figure 9: The process for loading models with customized GCN operators in [PITH_FULL_IMAGE:figures/full_fig_p005_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of on-meter forecasting results: GCN vs. GraphSAGE [PITH_FULL_IMAGE:figures/full_fig_p006_11.png] view at source ↗
read the original abstract

This paper presents a detailed study of how graph neural networks can be used on edge intelligent meters in a microgrid to forecast photovoltaic power generation. The problem background and the adopted technologies are introduced, including ONNX and ONNX Runtime. The hardware and software specifications of the smart meter are also briefly described. Then, the paper focuses on the training and deployment of two graph machine learning models, GCN and GraphSAGE, with particular emphasis on developing and deploying a customized ONNX operator for GCN. Finally, a case study is conducted using real datasets from a village microgrid. The performance of the two models is compared on both the PC and the smart meter, exhibiting successful deployments and executions on the smart meter.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. This paper presents a case study on applying graph neural networks (GCN and GraphSAGE) to photovoltaic power forecasting for smart meters in a village microgrid. It introduces ONNX and ONNX Runtime, describes the target smart meter hardware and software, details the training of the two models with emphasis on a custom ONNX operator developed for GCN, and reports a comparison of model performance when executed on both a PC and the smart meter, claiming successful on-device deployments and executions.

Significance. If the on-device results hold with acceptable latency, memory footprint, and minimal accuracy loss relative to PC execution, the work would provide a concrete demonstration of graph ML feasibility on constrained edge hardware for real-time grid-edge applications. The explicit development of a custom ONNX operator for GCN and the use of actual microgrid datasets are positive elements that could inform similar edge-intelligence efforts in energy systems.

major comments (1)
  1. The central claim that the models exhibit 'successful deployments and executions on the smart meter' is not supported by quantitative evidence. The manuscript should report concrete on-device metrics (inference latency, peak memory usage, and accuracy measures such as MAE/RMSE) for both GCN (with the custom operator) and GraphSAGE, together with direct deltas versus the PC baseline. Without these numbers the weakest assumption—that the custom operator runs efficiently with no major accuracy or speed penalty on the meter—remains untested and the deployment success assertion cannot be evaluated.
minor comments (1)
  1. The abstract summarizes the comparison and successful deployment but supplies no numerical performance figures or error statistics; adding at least one key metric (e.g., on-meter latency or accuracy) would make the abstract more informative.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive review and the opportunity to strengthen the quantitative support for our claims. We address the major comment below.

read point-by-point responses

  1. Referee: The central claim that the models exhibit 'successful deployments and executions on the smart meter' is not supported by quantitative evidence. The manuscript should report concrete on-device metrics (inference latency, peak memory usage, and accuracy measures such as MAE/RMSE) for both GCN (with the custom operator) and GraphSAGE, together with direct deltas versus the PC baseline. Without these numbers the weakest assumption—that the custom operator runs efficiently with no major accuracy or speed penalty on the meter—remains untested and the deployment success assertion cannot be evaluated.

    Authors: We agree that the manuscript would be strengthened by explicit quantitative metrics to support the on-device deployment claims. The revised manuscript will include a dedicated table (or expanded results section) reporting inference latency, peak memory usage, and accuracy (MAE/RMSE) for both the GCN model (with the custom ONNX operator) and GraphSAGE when executed on the smart meter. Direct comparisons and deltas versus the PC baseline will be added to quantify any differences in accuracy or speed. revision: yes

Circularity Check

0 steps flagged

No circularity: pure empirical case study with no derivations or self-referential predictions

full rationale

The paper describes an implementation and deployment case study for GCN and GraphSAGE models on smart meter hardware for PV power forecasting, including a custom ONNX operator. It reports training on real microgrid datasets and performance comparisons between PC and meter. No mathematical derivations, equations, fitted parameters presented as predictions, uniqueness theorems, or ansatzes are present. Central claims rest on direct empirical execution results rather than any chain that reduces to its own inputs by construction. This is a standard non-circular empirical report.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is an applied engineering case study with no mathematical derivations, free parameters, axioms, or postulated entities.

pith-pipeline@v0.9.0 · 5433 in / 917 out tokens · 30033 ms · 2026-05-10T16:48:34.274514+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

  1. [1]

    The core of this trend is ``on-device inference

    11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEcompsocconfonly \@IEEEauthordefaulttextstyle \@IEEEcompsocnotconfonly \@IEEEauthor...

    work page 2024

  2. [2]

    Embedded Machine Learning for Solar PV Power Regulation in a Remote Microgrid,

    Y. Zhu, L. Xu and J. Huang, "Embedded Machine Learning for Solar PV Power Regulation in a Remote Microgrid," 2024 7th International Conference on Power and Energy Applications (ICPEA), Taiyuan, China, 2024, pp. 756-761, doi: 10.1109/ICPEA63589.2024.10784815

    work page doi:10.1109/icpea63589.2024.10784815 2024

  3. [3]

    Flexible and Fully Quantized Lightweight TinyissimoYOLO for Ultra-Low-Power Edge Systems,

    J. Moosmann, H. Müller, N. Zimmerman, G. Rutishauser, L. Benini and M. Magno, "Flexible and Fully Quantized Lightweight TinyissimoYOLO for Ultra-Low-Power Edge Systems," in IEEE Access, vol. 12, pp. 75093-75107, 2024, doi: 10.1109/ACCESS.2024.3404878

    work page doi:10.1109/access.2024.3404878 2024

  4. [4]

    Edge Computing for Microgrid via MATLAB Embedded Coder and Low-Cost Smart Meters,

    L. Xu, J. Huang, S. Yang and Y. Zhu, "Edge Computing for Microgrid via MATLAB Embedded Coder and Low-Cost Smart Meters," 2024 8th International Conference on Smart Grid and Smart Cities (ICSGSC), Shanghai, China, 2024, pp. 234-239, doi: 10.1109/ICSGSC62639.2024.10813710

    work page doi:10.1109/icsgsc62639.2024.10813710 2024

  5. [5]

    DEST-GNN: A double-explored spatio-temporal graph neural network for multi-site intra-hour PV power forecasting,

    Y. Yang, Y. Liu, Y. Zhang, S. Shu, and J. Zheng, “DEST-GNN: A double-explored spatio-temporal graph neural network for multi-site intra-hour PV power forecasting,” Applied Energy, vol. 378, p. 124744, 2025, doi: 10.1016/j.apenergy.2024.124744

    work page doi:10.1016/j.apenergy.2024.124744 2025

  6. [6]

    Machine learning-driven intelligent voltage control in renewable energy grids,

    A. Murugesan, K. Durgadevi, D. Contractor, and Y. Rathod, “Machine learning-driven intelligent voltage control in renewable energy grids,” Electric Power Systems Research, vol. 248, p. 111869, 2025, doi: 10.1016/j.epsr.2025.111869

    work page doi:10.1016/j.epsr.2025.111869 2025

  7. [7]

    [Online]

    Open Neural Network Exchange, “Home,” ONNX. [Online]. Available: https://onnx.ai/

    work page

  8. [8]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in Proc. Int. Conf. Learn. Representations (ICLR), 2017. [Online]. Available: https://openreview.net/forum?id=SJU4ayYgl

    work page 2017

  9. [9]

    Inductive representation learning on large graphs,

    W. L. Hamilton, R. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” in Proc. 31st Int. Conf. Neural Inf. Process. Syst. (NeurIPS), Long Beach, CA, USA, 2017, pp. 1025–1035

    work page 2017

  10. [10]

    ONNX Runtime | Home,

    ONNX Runtime developers, “ONNX Runtime | Home,” ONNX Runtime. [Online]. Available: https://onnxruntime.ai/

    work page

  11. [11]

    TorchScript-based ONNX Exporter,

    PyTorch Contributors, “TorchScript-based ONNX Exporter,” PyTorch Documentation. [Online]. Available: https://docs.pytorch.org/docs/stable/onnx\_torchscript.html

    work page

  12. [12]

    Custom operators,

    ONNX Runtime, “Custom operators,” ONNX Runtime Docs. [Online]. Available: https://onnxruntime.ai/docs/reference/operators/add-custom-op.html

    work page