pith. sign in

arxiv: 2408.17366 · v1 · submitted 2024-08-30 · 💻 cs.LG · cs.AI

Leveraging Graph Neural Networks to Forecast Electricity Consumption

Eloi Campagne , Yvenn Amara-Ouali , Yannig Goude , Argyris Kalogeratos This is my paper

Pith reviewed 2026-05-23 21:28 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords graph neural networkselectricity consumption forecastinggraph convolutional networksgraph sagedecentralized energy networksgeneralized additive modelsdemand forecasting
0
0 comments X

The pith

Graph neural networks capture interconnections among regions to forecast electricity consumption beyond conventional models.

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

The paper establishes that modeling electricity consumption as graphs of regional nodes enables Graph Convolutional Networks and GraphSAGE to incorporate varying levels of interconnectedness and information sharing. This extends the standard Generalized Additive Model approach to handle the added complexity from renewable sources and decentralized networks. A sympathetic reader would care because better forecasts directly support grid stability amid rising uncertainty. The work introduces tailored methods to infer graphs from consumption data plus an evaluation framework covering performance and explainability. Experiments compare the approach on both synthetic data and real consumption across French mainland regions.

Core claim

The authors claim that graph representations of combined regional loads, processed through Graph Convolutional Networks or GraphSAGE, allow multiple degrees of node interconnectedness and information sharing that the conventional Generalized Additive Model framework does not provide, with new graph inference techniques from consumption data enabling this on both synthetic and French regional datasets.

What carries the argument

Graph inference methods from consumption data that define connections among regional load nodes for processing by Graph Convolutional Networks or GraphSAGE.

If this is right

  • Models gain the ability to incorporate multiple levels of interconnectedness among consumption nodes.
  • A dedicated evaluation framework measures both forecasting accuracy and explainability of the predictions.
  • The same graph construction and modeling steps apply equally to synthetic test cases and real French mainland regional data.

Where Pith is reading between the lines

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

  • The graph approach could reduce forecast errors during periods of high renewable penetration by better modeling cross-region effects.
  • Similar graph inference steps might transfer to forecasting tasks in other networked systems such as water distribution or transportation loads.
  • Explainability outputs from the models could reveal which regional connections most influence national demand predictions.

Load-bearing premise

The inferred graphs from consumption data accurately reflect the true spatial distribution and relational structure of the decentralized electricity network.

What would settle it

A side-by-side test on French regional consumption data where the graph neural network models show no accuracy gain over a Generalized Additive Model baseline would disprove the central claim.

Figures

Figures reproduced from arXiv: 2408.17366 by Argyris Kalogeratos, Eloi Campagne, Yannig Goude, Yvenn Amara-Ouali.

Figure 1
Figure 1. Figure 1: Example of a message passing layer in a GNN. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Graph corresponding to Wλ with λ = 0.71 and kernel bandwidth σ = 478.3. Dimension Reduction Statistical Transformation [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Inferring a graph from data using a dimension reduction algorithm and a statistical trans [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Generated temperature and load using pairwise influence between the regions ( [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Error variation by model on the synthetic test set. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Weights associated with the experts on the synthetic datasets. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Explanation graphs in June 2019 obtained from the [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Spline (dashed line) and predicted (scatter plot) effects. The distribution of the generated [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
read the original abstract

Accurate electricity demand forecasting is essential for several reasons, especially as the integration of renewable energy sources and the transition to a decentralized network paradigm introduce greater complexity and uncertainty. The proposed methodology leverages graph-based representations to effectively capture the spatial distribution and relational intricacies inherent in this decentralized network structure. This research work offers a novel approach that extends beyond the conventional Generalized Additive Model framework by considering models like Graph Convolutional Networks or Graph SAGE. These graph-based models enable the incorporation of various levels of interconnectedness and information sharing among nodes, where each node corresponds to the combined load (i.e. consumption) of a subset of consumers (e.g. the regions of a country). More specifically, we introduce a range of methods for inferring graphs tailored to consumption forecasting, along with a framework for evaluating the developed models in terms of both performance and explainability. We conduct experiments on electricity forecasting, in both a synthetic and a real framework considering the French mainland regions, and the performance and merits of our approach are discussed.

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

2 major / 2 minor

Summary. The manuscript claims that graph neural networks (GCN and GraphSAGE) extend electricity consumption forecasting beyond the Generalized Additive Model (GAM) framework by representing regional loads as nodes in graphs inferred from consumption data, thereby capturing spatial distribution and enabling information sharing. Tailored graph inference methods are introduced, and the approach is evaluated for performance and explainability on both synthetic data and real French mainland regional consumption data.

Significance. If the results hold after addressing validation gaps, the work could advance forecasting models for decentralized energy networks by providing a structured way to incorporate relational information, with the dual evaluation on performance and explainability as a constructive element. The introduction of consumption-tailored inference techniques is a clear strength that could be built upon in future studies.

major comments (2)
  1. [Abstract and Evaluation Framework] Abstract and described evaluation framework: no quantitative results, error bars, ablation details, or statistical tests are supplied to support the performance claims of GNN models over GAM baselines. This prevents verification of whether observed gains arise from graph-based interconnectedness rather than added model capacity.
  2. [Graph Inference Methods] Graph inference methods (as introduced for consumption forecasting): the central claim requires that inferred graphs meaningfully encode the spatial/relational structure of the decentralized electricity network, yet no external validation is reported against known transmission capacities, geographic distances, or inter-regional flows. Without this check, performance differences could be explained by regularization effects alone rather than the claimed information sharing.
minor comments (2)
  1. [Abstract] The abstract would benefit from including at least one key quantitative result (e.g., MAPE improvement on the French dataset) to allow readers to gauge the magnitude of the reported gains.
  2. [Methods] Clarify the precise definition of a 'node' (combined load of a subset of consumers) with an accompanying diagram or equation in the methods section to improve readability for readers unfamiliar with regional aggregation.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive comments, which highlight important aspects of result presentation and validation. We address each major comment below and indicate planned revisions to improve the manuscript.

read point-by-point responses

  1. Referee: [Abstract and Evaluation Framework] Abstract and described evaluation framework: no quantitative results, error bars, ablation details, or statistical tests are supplied to support the performance claims of GNN models over GAM baselines. This prevents verification of whether observed gains arise from graph-based interconnectedness rather than added model capacity.

    Authors: We agree that the abstract lacks quantitative support for the claims. In the revised manuscript, we will update the abstract to report key performance metrics (e.g., MAE or RMSE improvements with standard deviations across runs) from the synthetic and French regional experiments. We will also augment the evaluation framework section with explicit ablation studies that control for model capacity (such as non-graph neural baselines with comparable parameter counts) and include statistical tests (e.g., paired significance tests) to assess whether improvements are attributable to the graph structure rather than capacity alone. revision: yes

  2. Referee: [Graph Inference Methods] Graph inference methods (as introduced for consumption forecasting): the central claim requires that inferred graphs meaningfully encode the spatial/relational structure of the decentralized electricity network, yet no external validation is reported against known transmission capacities, geographic distances, or inter-regional flows. Without this check, performance differences could be explained by regularization effects alone rather than the claimed information sharing.

    Authors: The graph inference procedures are intentionally data-driven from consumption time series to recover latent relational patterns. We acknowledge the value of external validation. In revision we will incorporate available proxies such as geographic distances between regions and compare performance against distance-based graphs. The synthetic data experiments already provide controlled recovery of known graph structures. However, detailed transmission capacities and inter-regional flow data are not available in our public datasets, preventing comprehensive external checks. revision: partial

standing simulated objections not resolved
  • External validation of inferred graphs against transmission capacities and inter-regional flows, as such granular data is unavailable for the French regions in the study.

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper introduces graph inference methods and applies standard GNN architectures (GCN, GraphSAGE) for supervised forecasting on synthetic and French regional load data. No equations, derivations, or self-citations reduce the claimed performance gains or graph-based information sharing to fitted inputs by construction, self-definitions, or load-bearing prior results from the same authors. The methodology is self-contained via explicit training and evaluation against external benchmarks (synthetic data and real consumption records), with no renaming of known results or ansatz smuggling.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review; the central modeling choice is that consumption-derived graphs meaningfully encode spatial dependencies, which is treated as a domain assumption rather than derived.

axioms (1)
  • domain assumption Consumption time series contain sufficient signal to infer meaningful spatial graphs for forecasting
    The paper states it introduces methods for inferring graphs tailored to consumption forecasting without providing external validation of those graphs.

pith-pipeline@v0.9.0 · 5710 in / 1151 out tokens · 20130 ms · 2026-05-23T21:28:23.934882+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

43 extracted references · 43 canonical work pages

  1. [1]

    Agarwal, R., Melnick, L., Frosst, N., Zhang, X., Lengerich, B., Caruana, R., Hinton, G.: Neural Additive Models: Interpretable Machine Learning with Neural Nets (2021)

    work page 2021

  2. [2]

    Sustainability (2021)

    Alasali, F., Nusair, K., Alhmoud, L., Zarour, E.: Impact of the Covid-19 pandemic on electricity demand and load forecasting. Sustainability (2021)

    work page 2021

  3. [3]

    Antoniadis, A., Gaucher, S., Goude, Y.: Hierarchical transfer learning with applications for electricity load forecasting (2022)

    work page 2022

  4. [4]

    Journal of the Royal Statistical Society Series B: Statistical Methodology (2020)

    Apley, D.W., Zhu, J.: Visualizing the effects of predictor variables in black box supervised learning models. Journal of the Royal Statistical Society Series B: Statistical Methodology (2020)

    work page 2020

  5. [5]

    Interna- tional Journal of Forecasting (2022)

    Brégère, M., Huard, M.: Online hierarchical forecasting for power consumption data. Interna- tional Journal of Forecasting (2022)

    work page 2022

  6. [6]

    Cambridge university press (2006)

    Cesa-Bianchi, N., Lugosi, G.: Prediction, learning, and games. Cambridge university press (2006)

    work page 2006

  7. [7]

    IEEE Access (2020)

    Chen, K., Chen, F., Lai, B., Jin, Z., Liu, Y., Li, K., Wei, L., Wang, P., Tang, Y., Huang, J., et al.: Dynamic spatio-temporal graph-based cnns for traffic flow prediction. IEEE Access (2020)

    work page 2020

  8. [8]

    Distill (2021)

    Daigavane, A., Ravindran, B., Aggarwal, G.: Understanding convolutions on graphs. Distill (2021)

    work page 2021

  9. [9]

    Doumèche, N., Allioux, Y., Goude, Y., Rubrichi, S.: Human spatial dynamics for electricity demand forecasting: the case of france during the 2022 energy crisis (2023)

    work page 2022

  10. [10]

    IEEE (2010)

    Fan, S., Hyndman, R.J.: Forecast short-term electricity demand using semi-parametric additive model. IEEE (2010)

    work page 2010

  11. [11]

    Journal of the American Statistical Association (2021)

    Fasiolo, M., Wood, S.N., Zaffran, M., Nedellec, R., Goude, Y.: Fast calibrated additive quantile regression. Journal of the American Statistical Association (2021)

    work page 2021

  12. [12]

    Feik, M., Lerch, S., Stühmer, J.: Graph Neural Networks and Spatial Information Learning for Post-Processing Ensemble Weather Forecasts (2024)

    work page 2024

  13. [13]

    International Journal of forecasting (2016)

    Gaillard,P.,Goude,Y.,Nedellec,R.:AdditivemodelsandrobustaggregationforGEFCom2014 probabilistic electric load and electricity price forecasting. International Journal of forecasting (2016)

    work page 2016

  14. [14]

    In: Conference on Learning Theory

    Gaillard, P., Stoltz, G., Van Erven, T.: A second-order bound with excess losses. In: Conference on Learning Theory. PMLR (2014)

    work page 2014

  15. [15]

    Garreau, D., Jitkrittum, W., Kanagawa, M.: Large sample analysis of the median heuristic (2018)

    work page 2018

  16. [16]

    Sustainable Energy, Grids and Networks (2023)

    Gilbert, C., Browell, J., Stephen, B.: Probabilistic load forecasting for the low voltage network: Forecast fusion and daily peaks. Sustainable Energy, Grids and Networks (2023)

    work page 2023

  17. [17]

    In: Proceedings of the AAAI conference on artificial intelligence (2019)

    Guo, S., Lin, Y., Feng, N., Song, C., Wan, H.: Attention based spatial-temporal graph convolu- tional networks for traffic flow forecasting. In: Proceedings of the AAAI conference on artificial intelligence (2019)

    work page 2019

  18. [18]

    Hamilton, W.L., Ying, R., Leskovec, J.: Inductive Representation Learning on Large Graphs (2018)

    work page 2018

  19. [19]

    (1986) Energy Forecasting with GNNs 17

    Hastie, T., Tibshirani, R.: Generalized additive models (with discussion), Statistical Science. (1986) Energy Forecasting with GNNs 17

    work page 1986

  20. [20]

    The Journal of Machine Learning Research (2021)

    Humbert, P., Le Bars, B., Oudre, L., Kalogeratos, A., Vayatis, N.: Learning laplacian ma- trix from graph signals with sparse spectral representation. The Journal of Machine Learning Research (2021)

    work page 2021

  21. [21]

    Foresight: The International Journal of Applied Forecasting (2006)

    Hyndman, R.J., et al.: Another look at forecast-accuracy metrics for intermittent demand. Foresight: The International Journal of Applied Forecasting (2006)

    work page 2006

  22. [22]

    Technological Forecasting and Social Change (2024)

    Jiang, H., Dong, Y., Dong, Y., Wang, J.: Power load forecasting based on spatial–temporal fusion graph convolution network. Technological Forecasting and Social Change (2024)

    work page 2024

  23. [23]

    Jiang,J.,Han,C.,Wang,J.:Buaa_bigscity:Spatial-TemporalGraphNeuralNetworkforWind Power Forecasting in Baidu KDD CUP 2022 (2023)

    work page 2022

  24. [24]

    Applied energy

    Jiang, P., Van Fan, Y., Klemeš, J.J.: Impacts of Covid-19 on energy demand and consumption: Challenges, lessons and emerging opportunities. Applied energy

    work page

  25. [25]

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization (2014)

    work page 2014

  26. [26]

    Kipf, T.N., Welling, M.: Semi-Supervised Classification with Graph Convolutional Networks (2017)

    work page 2017

  27. [27]

    The handbook of brain theory and neural networks (1995)

    LeCun, Y., Bengio, Y., et al.: Convolutional networks for images, speech, and time series. The handbook of brain theory and neural networks (1995)

    work page 1995

  28. [28]

    Long, M., Cao, Y., Wang, J., Jordan, M.I.: Learning Transferable Features with Deep Adap- tation Networks (2015)

    work page 2015

  29. [29]

    In: Annual conference of the IEEE industrial electronics society (2016)

    Marino, D.L., Amarasinghe, K., Manic, M.: Building energy load forecasting using deep neural networks. In: Annual conference of the IEEE industrial electronics society (2016)

    work page 2016

  30. [30]

    Massaoudi, M., Abu-Rub, H., Refaat, S.S., Chihi, I., Oueslati, F.S.: Deep Learning in Smart GridTechnology:AReviewofRecentAdvancementsandFutureProspects.IEEEAccess(2021)

    work page 2021

  31. [31]

    IEEE Transactions on Power Systems (2021)

    Obst, D., de Vilmarest, J., Goude, Y.: Adaptive Methods for Short-Term Electricity Load Forecasting During Covid-19 Lockdown in France. IEEE Transactions on Power Systems (2021)

    work page 2021

  32. [32]

    Proceedings of ISAP power (2011)

    Pierrot, A., Goude, Y.: Short-term electricity load forecasting with generalized additive models. Proceedings of ISAP power (2011)

    work page 2011

  33. [33]

    Rossi, E., Chamberlain, B., Frasca, F., Eynard, D., Monti, F., Bronstein, M.: Temporal graph networks for deep learning on dynamic graphs (2020)

    work page 2020

  34. [34]

    In: KDD workshop on mining temporal and sequential data (2004)

    Salvador, S., Chan, P.: FastDTW: Toward accurate dynamic time warping in linear time and space. In: KDD workshop on mining temporal and sequential data (2004)

    work page 2004

  35. [35]

    IEEE Transactions on Smart Grid (2017)

    Shi, H., Xu, M., Li, R.: Deep learning for household load forecasting—A novel pooling deep RNN. IEEE Transactions on Smart Grid (2017)

    work page 2017

  36. [36]

    Thielmann, A., Kruse, R.M., Kneib, T., Säfken, B.: Neural Additive Models for Location Scale and Shape: A Framework for Interpretable Neural Regression Beyond the Mean (2024)

    work page 2024

  37. [37]

    IEEE Transactions on Power Systems (2023)

    de Vilmarest, J., Browell, J., Fasiolo, M., Goude, Y., Wintenberger, O.: Adaptive probabilistic forecasting of electricity (net-)load. IEEE Transactions on Power Systems (2023)

    work page 2023

  38. [38]

    de Vilmarest, J., Goude, Y.: State-Space Models Win the IEEE DataPort Competition on Post-covid Day-ahead Electricity Load Forecasting (2021)

    work page 2021

  39. [39]

    Energy and Built Environment (2020)

    Williams, S., Short, M.: Electricity demand forecasting for decentralised energy management. Energy and Built Environment (2020)

    work page 2020

  40. [40]

    CRC press (2017)

    Wood, S.N.: Generalized additive models: an introduction with R. CRC press (2017)

    work page 2017

  41. [41]

    Ecological modelling (2002)

    Wood, S.N., Augustin, N.H.: GAMs with integrated model selection using penalized regression splines and applications to environmental modelling. Ecological modelling (2002)

    work page 2002

  42. [42]

    Journal of the American Statistical Association (2016)

    Wood, S.N., Pya, N., Säfken, B.: Smoothing parameter and model selection for general smooth models. Journal of the American Statistical Association (2016)

    work page 2016

  43. [43]

    Advances in neural information processing systems (2019)

    Ying, Z., Bourgeois, D., You, J., Zitnik, M., Leskovec, J.: GNNExplainer: Generating explana- tions for graph neural networks. Advances in neural information processing systems (2019)

    work page 2019