Pith. sign in

REVIEW 3 major objections 5 minor 34 references

Spatiotemporal Graph Neural Networks in short term load forecasting: Does adding Graph Structure in Consumption Data Improve Predictions?

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Adding graph structure to household consumption data improves day-ahead forecasts for individual homes but worsens the aggregate forecast.

desk verdict Useful benchmark but the title overclaims: the residential gains come from attention-based arbitrary-topology models, not graph structure mined from consumption data. read the letter →

arxiv 2502.12175 v1 pith:SCKAETAY submitted 2025-02-14 cs.LG cs.AI

Quoc Viet Nguyen , Joaquin Delgado Fernandez , Sergio Potenciano Menci This is my paper · ORCID
classification cs.LGcs.AI
keywords short-termloadforecastingspatiotemporalgraphneuralnetworkssmartmeterdataconstructionresidentialaggregateday-aheadbenchmarkevaluation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper asks whether adding a graph structure to household consumption data improves short-term load forecasts, and its answer is a split verdict: graph features help at the individual household level, but not at the aggregate level. It reviews how STGNNs form graphs and process time and space, then benchmarks seven graph models against seasonal naive, VAR, GRU, and Transformer baselines on a 228-household subset of the Low Carbon London smart-meter data with a day-ahead horizon. The best graph models beat the temporal-only deep baselines on residential MAE and RMSE, yet summing those household forecasts gives errors worse than the simple seasonal baseline. If the result holds, practitioners should expect spatial information to improve per-household forecasts, not system-level totals.

What carries the argument

The load-bearing mechanism is message passing over a graph of households: each smart-meter node aggregates weighted information from its neighbours and updates its representation, so consumption at one house becomes an input to the forecast of another. Within that mechanism the paper distinguishes predefined graphs (similarity by Euclidean distance, DTW, correlation, or correntropy) from learnable graphs, and Time-then-Space architecture (temporal encoding followed by graph convolution) from Time-and-Space architecture (graph convolution interleaved with a recurrent cell). A second mechanism, the left-skewed error distribution at peak hours, is what the paper uses to explain why graph structure helps per household but not in aggregate: spatial propagation of underestimated spikes prevents errors from cancelling when forecasts are summed.

What would settle it

Run the same benchmark on a second smart-meter dataset, or on the full Low Carbon London sample rather than the 228-household subset, with more than one forecast horizon, and report confidence intervals or significance tests for the MAE and RMSE differences between graph models and GRU: if the residential improvement falls inside sampling noise in most splits, or if an aggregate graph forecast beats SeasonalNaive, the central claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is a conditional one: for day-ahead residential load forecasting on 228 households from a single sociodemographic group, adding spatial structure through a graph improves accuracy relative to temporal-only deep models. The best graph-based models—GCGRU, FC-GNN, and BP-GNN—achieve the lowest MAE and RMSE on most splits, while fully connected and bipartite topologies outperform signal-derived graphs. At the aggregate level, however, the graph models do not help: when the same household forecasts are summed, the seasonal naive baseline beats every deep model. The paper explains the discrepancy by the distribution of peak-hour errors: deep models, including graph models, underestimate spikes, and spatial propagation can spread rather than cancel those errors.

Load-bearing premise

The central claim assumes that performance differences measured on a single open dataset of 228 households from one socioeconomic group, with one forecasting horizon, one hyperparameter tuning pass, and no significance testing, are stable enough to show that graph structure helps generally.

Editorial extensions

If this is right

  • Residential-level forecasters should expect graph-based models such as GCGRU, FC-GNN, and BP-GNN to beat temporal-only deep models like GRU and Transformer on day-ahead MAE and RMSE.
  • System-level forecasters should not expect graph structure to help, because summed household forecasts from STGNNs are worse than a seasonal naive baseline in this study.
  • The choice of graph formation method is not decisive: simple fully connected and bipartite attention topologies perform as well as or better than graphs built from time-series similarity.
  • Learnable graph models such as AGCRN and GraphWaveNet do not consistently beat predefined-graph models and can overfit when the test month is farther from training.
  • Bipartite graphs offer a scalable alternative to fully connected graphs, since their interaction cost is linear in the number of virtual nodes rather than quadratic in households.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the winning graph topologies are not based on measured similarity, the useful signal may be latent consumer segmentation rather than pairwise correlation; building graphs from sociodemographic clusters or learned cluster assignments is a natural untested extension.
  • The aggregate-level failure suggests a bias, not just variance: deep models underestimate peak-hour consumption, and summing household predictions concentrates that bias. Adding a peak-hour residual correction or training on an asymmetric loss might let graph models beat the seasonal baseline in aggregate.
  • The benchmark covers only one day-ahead horizon, one socioeconomic subgroup, and one dataset, so the split verdict may not transfer to other horizons, mixed consumer groups, or data with missing values; multi-horizon and cross-dataset replication is the direct next step.
  • If graph structure helps only through attention-like weighting of other households, then a simpler temporal model with cross-household attention and no explicit graph constraints might capture the same residential gains at lower cost.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper surveys spatiotemporal graph neural networks (STGNNs) for short-term load forecasting and benchmarks seven STGNN variants against SeasonalNaive, VAR, GRU, and Transformer on 228 households from the Low Carbon London dataset. The benchmark uses three train/validation/test splits at the residential and aggregate levels, with MAE, MAPE, and RMSE reported as means over five runs. The main reported findings are that graph-based models can improve residential forecasting compared with temporal-only baselines, that simple full or bipartite graph topologies with attention often outperform graphs derived from consumption signals, and that aggregated residential forecasts are worse than a seasonal naive baseline at the aggregate level.

Significance. If the residential-level result is taken at face value, the paper provides a useful reproducible comparison for practitioners choosing among STGNN architectures. The survey component is informative, and the benchmark includes multiple splits, an open dataset, and a public code repository. However, the significance is weakened by the conflation of graph structure with attention-based arbitrary topologies, the absence of significance tests for small differences, and the indirect nature of the aggregate-level experiment. The strongest contribution is the careful documentation of an empirical comparison, not an established causal claim about graph structure.

major comments (3)
  1. [Section 4.2, Table 3] The paper's headline claim that adding graph structure extracted from consumption data improves residential forecasting is not cleanly supported by the reported results. The largest and most consistent gains over the GRU baseline come from FC-GNN (e.g., Split 3 MAE 146.9 vs 153.1) and BP-GNN (148.0), whose topologies are fully connected or bipartite and whose edge weights are generated by attention from the input; they use no graph structure mined from consumption data. The models that do use predefined or learnable graphs from consumption signals show smaller or inconsistent gains (e.g., GraphWavenet is worse than GRU in Split 2, 128.4 vs 126.5, and Split 3, 155.9 vs 153.1). The paper itself concedes in Section 4.2 that the success of these 'naive' topologies 'questions if the graph formation based on signals or learnable parameters is effective.' Consequently, the observed benefit cannot be attributed to the spatial structure of consumption data; it may reflect the attention mechanism or added model capacity. This conflation is load-bearing because the title and abstract ask whether graph structure in consumption data improves predictions.
  2. [Section 4.3, Table 4] The aggregate-level conclusion is based on summing the per-household forecasts of each residential model rather than training or evaluating models directly on the aggregate load series. Table 4 therefore measures the aggregation error of residential forecasts, not the performance of STGNNs as aggregate-level forecasting models. The abstract's statement that the graph benefit 'is not reflected at the aggregate level' overstates what is tested; a model trained on the aggregate series could in principle behave differently. The authors should either reframe this result as 'aggregating residential forecasts' or add a direct aggregate-level experiment before drawing conclusions about aggregate forecasting performance.
  3. [Section 4, Tables 2–4] All headline comparisons report only means and standard deviations over five runs, with no significance tests, confidence intervals, or effect sizes. Several differences that the text interprets as improvement are smaller than the reported variability. For example, in Split 1 the GRU MAE is 89.5(0.1) while GRUGCN, T-GCN, and GCGRU are 89.0(0.2), 88.9(0.5), and 88.2(0.2); with n=5 and overlapping spreads, these differences are not demonstrably reliable. Similarly, in Table 2 the graph-formation comparisons for GCGRU (MAE range 149.0–149.6 with standard deviations up to 0.5) are within noise. Because the paper's main positive claim rests on small differences, the authors should add paired significance tests or show that the differences are consistent across the five runs (e.g., per-run win rates).
minor comments (5)
  1. [Footnote 1] The footnote contains a typo: 'pen access' should read 'open access', and the sentence should be checked for grammatical completeness.
  2. [Section 3.2, Eq. (1) and Section 3.3, Eq. (2)] Eq. (1) defines MAE as a sum without normalization, while Eq. (2) defines MAE as a mean over N and T; please clarify whether the training loss is the mean or the sum, and fix the index inconsistency between i, n, and t.
  3. [Table 2] The column header appears as 'ModelGraph formation'; this is likely a formatting error and should be corrected for readability.
  4. [Figure 4] The text describes the x-axis as errors, the y-axis as models, and the z-axis as frequency, but the resulting 'histogram' with models on a categorical axis is difficult to interpret; a conventional boxplot or violin plot of the per-household errors would be clearer.
  5. [Section 3.1] The phrase 'consistent with different time scales' is misleading because the three splits vary the training period and test month, not the forecast horizon or data sampling resolution; consider rewording as 'different training and test periods.'

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the benchmark compares held-out STGNN forecasts against independent temporal baselines, and no derivation reduces to its own inputs.

full rationale

The paper is an empirical benchmark, not a formal derivation. The central claim—that adding graph features can improve residential load forecasts—rests on held-out test errors in Table 3 compared against temporal baselines (GRU, Transformer, VAR, SeasonalNaive). No equation defines an output in terms of the claimed result, no fitted parameter is renamed as a prediction, and no load-bearing premise is imported from the authors' prior work. The only self-citations ([15], [16]) appear as references for Pearson- and DTW-based graph-construction heuristics; they are not used to justify the paper's conclusions. The acknowledged limitations (single day-ahead horizon, no exogenous variables, a selected model set) weaken generalizability but are not circular. The paper itself concedes that 'naive' topologies such as fully connected and bipartite graphs outperform signal-based graphs, which undercuts a strong causal reading of the title but does not make the argument circular. Hyperparameter tuning on validation data and selecting the best graph-formation variant on Split 3 are methodological concerns, not instances of a prediction reducing to its input by construction. Therefore the analysis finds no significant circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim depends on the experimental setup, so the main ledger items are hyperparameters selected on validation data and implicit assumptions about the representativeness of the dataset and the validity of the graph-based modeling approach.

free parameters (5)
  • learning rate = tuned per model, not listed
    Tuned on validation data for each model; affects the comparison.
  • batch size = tuned, not listed
    Tuned on validation data; affects training dynamics.
  • training window W = tuned, not listed
    The length of historical input is tuned on validation data.
  • graph similarity threshold or edge construction knobs = not disclosed
    The paper states edges exist when similarity surpasses a threshold, but the threshold value is not reported for the predefined graphs.
  • number of virtual nodes K in BP-GNN = not reported
    This hyperparameter controls the bipartite graph size and is not specified.
assumptions (4)
  • domain assumption Message passing on a graph built from consumption similarity captures useful spatial dependencies for load forecasting.
    Section 2.2 assumes that aggregating neighboring consumption improves prediction, which is exactly what the benchmark tests.
  • domain assumption The selected 228 households from one Acorn group represent general residential consumption.
    Section 3.1 restricts the data to one socio-demographic group, limiting generalizability.
  • domain assumption Standard implementations from the tsl package are correct and comparable across models.
    Section 3.2 relies on the external tsl library without verifying implementation details.
  • standard math VAR baseline assumptions of stationarity and invertibility hold approximately for the data.
    Section 3.3 introduces VAR as a statistical baseline with standard time series assumptions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spatiotemporal Graph Neural Networks in short term load forecasting: Does adding Graph Structure in Consumption Data Improve Predictions?." pith.science (2026). https://pith.science/paper/SCKAETAY

@misc{pith2026250212175,
  author       = {Pith},
  title        = {Pith review of: Spatiotemporal Graph Neural Networks in short term load forecasting: Does adding Graph Structure in Consumption Data Improve Predictions?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SCKAETAY}},
  note         = {Machine review of arXiv:2502.12175}
}
read the original abstract

Short term Load Forecasting (STLF) plays an important role in traditional and modern power systems. Most STLF models predominantly exploit temporal dependencies from historical data to predict future consumption. Nowadays, with the widespread deployment of smart meters, their data can contain spatiotemporal dependencies. In particular, their consumption data is not only correlated to historical values but also to the values of neighboring smart meters. This new characteristic motivates researchers to explore and experiment with new models that can effectively integrate spatiotemporal interrelations to increase forecasting performance. Spatiotemporal Graph Neural Networks (STGNNs) can leverage such interrelations by modeling relationships between smart meters as a graph and using these relationships as additional features to predict future energy consumption. While extensively studied in other spatiotemporal forecasting domains such as traffic, environments, or renewable energy generation, their application to load forecasting remains relatively unexplored, particularly in scenarios where the graph structure is not inherently available. This paper overviews the current literature focusing on STGNNs with application in STLF. Additionally, from a technical perspective, it also benchmarks selected STGNN models for STLF at the residential and aggregate levels. The results indicate that incorporating graph features can improve forecasting accuracy at the residential level; however, this effect is not reflected at the aggregate level

Figures

Figures reproduced from arXiv: 2502.12175 by the authors.

Figure 2
Figure 2. T&S architecture. 2.4 Examples of STGNN models Given the different architectures of STGNN models, we review existing models that are representatives of the described architectures. We present and compare the similarity between models based on the components and architectures in Section 2.2 and 2.3. GRUGCN [22] This model is of type TTS. It uses the Gated Recurrent Unit (GRU) (a variant of RNN) as a temporal processi… view at source ↗
Figure 3
Figure 3. Train-validation-test split settings. Specifically, the training periods for splits 1, 2, and 3 span from January 1, 2013, to the day before July 1, September 1, and November 1, respectively. The validation and testing periods of each split cover the month immediately after the training period. The partition aims to see if the performance comparison is consistent with different time scales. 3.2 Model training For al… view at source ↗
Figure 4
Figure 4. Distribution of differences between forecasts and ground truth. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 29 canonical work pages

  1. [1]

    Manage- ment of an academic hpc research computing facility: The ulhpc experience 2.0,

    S. V ARRETTE, H. CARTIAUX, S. PETER, E. KIEFFER, T. V ALETTE, and A. OLLOH, “Manage- ment of an academic hpc research computing facility: The ulhpc experience 2.0,” in 6th High Perfor- mance Computing and Cluster Technologies Conference (HPCCT 2022) . Association for Computing Machinery (ACM), July 2022

    work page 2022

  2. [2]

    A comprehensive review on deep learning approaches for short-term load forecasting,

    Y. Eren and ˙I. K¨ u¸ c¨ ukdemiral, “A comprehensive review on deep learning approaches for short-term load forecasting,” Renewable and Sustainable Energy Reviews , vol. 189, p. 114031, 2024

    work page 2024

  3. [3]

    Big data analytics in smart grids: a review,

    Y. Zhang, T. Huang, and E. Bompard, “Big data analytics in smart grids: a review,” Energy Informat- ics, vol. 1, 08 2018

    work page 2018

  4. [4]

    Short-term residential load forecasting using Graph Convolutional Recurrent Neural Networks,

    S. Arastehfar, M. Matinkia, and M. R. Jabbarpour, “Short-term residential load forecasting using Graph Convolutional Recurrent Neural Networks,” Engineering Applications of Artificial Intelligence , vol. 116, p. 105358, Nov. 2022. [Online]. Available: https://www.sciencedirect.com/science/article/pii/ S0952197622003761

    work page 2022

  5. [5]

    Probabilistic electric load forecasting: A tutorial review,

    T. Hong and S. Fan, “Probabilistic electric load forecasting: A tutorial review,” International Journal of Forecasting , vol. 32, no. 3, pp. 914–938, Jul. 2016. [Online]. Available: https: //linkinghub.elsevier.com/retrieve/pii/S0169207015001508

    work page 2016

  6. [6]

    Time-series forecasting models for smart meters data: An empirical comparison and analysis,

    I. Moustati, N. Gherabi, and M. Saadi, “Time-series forecasting models for smart meters data: An empirical comparison and analysis,” J. Eur. Syst. Autom. , vol. 57, no. 05, pp. 1419–1427, Oct. 2024

    work page 2024

  7. [7]

    Understanding multi-scale spatiotemporal energy consumption data: A visual analysis approach,

    J. Wu, Z. Niu, X. Li, L. Huang, P. S. Nielsen, and X. Liu, “Understanding multi-scale spatiotemporal energy consumption data: A visual analysis approach,” Energy, vol. 263, p. 125939, 2023. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0360544222028250

    work page 2023

  8. [8]

    Spatial-temporal graph neural network for traffic forecasting: An overview and open research issues,

    K.-H. N. Bui, J. Cho, and H. Yi, “Spatial-temporal graph neural network for traffic forecasting: An overview and open research issues,” Applied Intelligence , vol. 52, no. 3, pp. 2763–2774, Feb. 2022. [Online]. Available: https://doi.org/10.1007/s10489-021-02587-w

    work page doi:10.1007/s10489-021-02587-w 2022

Show all 34 references
  1. [9]

    Residential Electric Load Forecasting via Attentive Transfer of Graph Neural Networks,

    W. Lin and D. Wu, “Residential Electric Load Forecasting via Attentive Transfer of Graph Neural Networks,” in Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence . Montreal, Canada: International Joint Conferences on Artificial Intelligence O...

    2021

  2. [10]

    Stgnet: Short-term residential load forecasting with spatial–temporal gated fusion network,

    D. Feng, D. ao Li, Y. Zhou, J. Zhao, and K. Zhang, “Stgnet: Short-term residential load forecasting with spatial–temporal gated fusion network,” Energy Science & Engineering , vol. 12, pp. 541 – 560,

  3. [11]

    Graph construction on complex spatiotemporal data for enhancing graph neural network-based approaches,

    S. Bloemheuvel, J. van den Hoogen, and M. Atzmueller, “Graph construction on complex spatiotemporal data for enhancing graph neural network-based approaches,” International Journal of Data Science and Analytics , vol. 18, no. 2, pp. 157–174, Aug. 2024. [Online]. Available: htt...

    2024 doi

  4. [12]

    Short-term load forecasting using spatial-temporal embedding graph neural network,

    C. Wei, D. Pi, M. Ping, and H. Zhang, “Short-term load forecasting using spatial-temporal embedding graph neural network,” Electric Power Systems Research , vol. 225, p. 109873, Dec. 2023. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0378779623007617

    2023

  5. [13]

    Graph Deep Learning for Time Series Forecasting,

    A. Cini, I. Marisca, D. Zambon, and C. Alippi, “Graph Deep Learning for Time Series Forecasting,” Oct. 2023, arXiv:2310.15978. [Online]. Available: http://arxiv.org/abs/2310.15978

    2023 arXiv

  6. [14]

    A short-term residential load forecasting scheme based on the multiple correlation-temporal graph neural networks,

    Y. Wang, L. Rui, J. Ma, and Q. jin, “A short-term residential load forecasting scheme based on the multiple correlation-temporal graph neural networks,” Applied Soft Computing , vol. 146, p. 110629, Oct. 2023. [Online]. Available: https://www.sciencedirect.com/science/article/...

    2023

  7. [15]

    Privacy-preserving federated learning for residential short-term load forecasting,

    J. D. Fern´ andez, S. P. Menci, C. M. Lee, A. Rieger, and G. Fridgen, “Privacy-preserving federated learning for residential short-term load forecasting,” Applied Energy, vol. 326, p. 119915, Nov. 2022. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S03...

    2022

  8. [16]

    Towards a peer-to-peer residential short-term load fore- casting with federated learning,

    J. D. Fernandez, S. P. Menci, and I. Pavic, “Towards a peer-to-peer residential short-term load fore- casting with federated learning,” in 2023 IEEE Belgrade PowerTech , 2023, pp. 1–6

    2023

  9. [17]

    Graph-based Time Series Clustering for End-to-End Hierarchical Forecasting,

    A. Cini, D. Mandic, and C. Alippi, “Graph-based Time Series Clustering for End-to-End Hierarchical Forecasting,” May 2023, arXiv:2305.19183 [cs]. [Online]. Available: http://arxiv.org/abs/2305.19183

    2023 arXiv

  10. [18]

    Graph WaveNet for Deep Spatial-Temporal Graph Modeling

    Z. Wu, S. Pan, G. Long, J. Jiang, and C. Zhang, “Graph WaveNet for Deep Spatial-Temporal Graph Modeling.”

  11. [19]

    Sequence to Sequence Learning with Neural Networks,

    I. Sutskever, O. Vinyals, and Q. V. Le, “Sequence to Sequence Learning with Neural Networks,” Dec. 2014, arXiv:1409.3215 [cs]

    2014 arXiv

  12. [20]

    Short-Term Load Forecasting of Electricity Demand for the Residential Sector Based on Modelling Techniques: A Systematic Review,

    F. Rodrigues, C. Cardeira, J. M. F. Calado, and R. Melicio, “Short-Term Load Forecasting of Electricity Demand for the Residential Sector Based on Modelling Techniques: A Systematic Review,” Energies, vol. 16, no. 10, p. 4098, Jan. 2023, number: 10 Publisher: Multidisciplinary...

    2023

  13. [21]

    Spatio-Temporal Short Term Load Forecasting Using Graph Neural Networks,

    H. Mansoor, M. Shabbir, M. Y. Ali, H. Rauf, M. Khalid, and N. Arshad, “Spatio-Temporal Short Term Load Forecasting Using Graph Neural Networks,” in 2023 12th International Conference on Renewable Energy Research and Applications (ICRERA) . Oshawa, ON, Canada: IEEE, Aug. 2023, ...

    2023

  14. [22]

    On the Equivalence Between Temporal and Static Equivariant Graph Representations,

    J. Gao and B. Ribeiro, “On the Equivalence Between Temporal and Static Equivariant Graph Representations,” in Proceedings of the 39th International Conference on Machine Learning. PMLR, Jun. 2022, pp. 7052–7076, iSSN: 2640-3498. [Online]. Available: https: //proceedings.mlr.pr...

    2022

  15. [23]

    Gated spatial-temporal graph neural network based short-term load forecasting for wide-area multiple buses,

    N. Huang, S. Wang, R. Wang, G. Cai, Y. Liu, and Q. Dai, “Gated spatial-temporal graph neural network based short-term load forecasting for wide-area multiple buses,” International Journal of Electrical Power & Energy Systems , vol. 145, p. 108651, Feb. 2023. [Online]. Availabl...

    2023

  16. [24]

    Adaptive Graph Convolutional Recurrent Network for Traffic Forecasting,

    L. BAI, L. Yao, C. Li, X. Wang, and C. Wang, “Adaptive Graph Convolutional Recurrent Network for Traffic Forecasting,” in Advances in Neural Information Processing Systems , vol. 33. Curran Associates, Inc., 2020, pp. 17 804–17 815. [Online]. Available: https: //proceedings.ne...

    2020

  17. [25]

    Spatial-Temporal Residential Short-Term Load Forecasting via Graph Neural Networks,

    W. Lin, D. Wu, and B. Boulet, “Spatial-Temporal Residential Short-Term Load Forecasting via Graph Neural Networks,” IEEE Transactions on Smart Grid , vol. 12, no. 6, pp. 5373– 5384, Nov. 2021, conference Name: IEEE Transactions on Smart Grid. [Online]. Available: https://ieeex...

    2021

  18. [26]

    Multivariate Time Series Forecasting with Latent Graph Inference,

    V. G. Satorras, S. S. Rangapuram, and T. Januschowski, “Multivariate Time Series Forecasting with Latent Graph Inference,” Mar. 2022, arXiv:2203.03423. [Online]. Available: http: //arxiv.org/abs/2203.03423

    2022 arXiv

  19. [27]

    Low Carbon London Project: Data from the Dynamic Time-of-Use Electricity Pricing Trial, 2013,

    G. Strbac, S. Tindemans, M. Woolf, M. Bilton, R. Carmichael, and J. R. Schofield, “Low Carbon London Project: Data from the Dynamic Time-of-Use Electricity Pricing Trial, 2013,” 2024. [Online]. Available: https://beta.ukdataservice.ac.uk/datacatalogue/doi/?id=7857#2

    2013

  20. [28]

    The acorn user guide,

    CACI, “The acorn user guide,” CACI, Tech. Rep., 2014

    2014

  21. [29]

    Torch Spatiotemporal,

    A. Cini and I. Marisca, “Torch Spatiotemporal,” 3 2022. [Online]. Available: https: //github.com/TorchSpatiotemporal/tsl

    2022

  22. [30]

    Attention Is All You Need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention Is All You Need,” Aug. 2023, arXiv:1706.03762. [Online]. Available: http://arxiv.org/abs/1706.03762

    2023 arXiv

  23. [31]

    Spatial and Temporal Attention-Enabled Transformer Network for Multivariate Short-Term Residential Load Forecasting,

    H. Zhao, Y. Wu, L. Ma, and S. Pan, “Spatial and Temporal Attention-Enabled Transformer Network for Multivariate Short-Term Residential Load Forecasting,” IEEE Transactions on Instrumentation and Measurement, vol. 72, pp. 1–11, 2023, conference Name: IEEE Transactions on Instru...

    2023

  24. [32]

    Unlocking the Potential of Deep Learning in Peak-Hour Series Forecasting,

    Z. Zhang, X. Wang, J. Xie, H. Zhang, and Y. Gu, “Unlocking the Potential of Deep Learning in Peak-Hour Series Forecasting,” in Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , Oct. 2023, pp. 4415–4419, arXiv:2307.01597 [cs]. [Onlin...

    2023 arXiv

  25. [33]

    ChatGPT,

    “ChatGPT,” https://chatgpt.com/. 13

  26. [2023]

    Available: https://api.semanticscholar.org/CorpusID:266077492

    [Online]. Available: https://api.semanticscholar.org/CorpusID:266077492

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.