TriForecaster: A Mixture of Experts Framework for Multi-Region Electric Load Forecasting with Tri-dimensional Specialization

Liang Sun; Linxiao Yang; Peiyuan Liu; Qiming Chen; Weiqi Chen; Zhaoyang Zhu; Zhipeng Zeng

REVIEW 4 major objections 6 minor 49 references

TriForecaster: A Mixture of Experts Framework for Multi-Region Electric Load Forecasting with Tri-dimensional Specialization

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A mixture-of-experts framework that specializes electric load forecasts by region, context, and time of day reduces average forecast error by 22.4% on four real-world multi-region datasets.

desk verdict An honest, well-structured applied MoE paper with a real deployment and plausible 22.4% gains, but the hard-sampling fusion is underspecified — the headline result rests on an undocumented training detail. read the letter →

arxiv 2508.09753 v1 pith:6IWVBG5Y submitted 2025-08-13 cs.LG

Zhaoyang Zhu , Zhipeng Zeng , Qiming Chen , Linxiao Yang , Peiyuan Liu , Weiqi Chen , Liang Sun This is my paper · ORCID

classification cs.LG

keywords electricloadforecastingmulti-regionmixtureofexpertsmulti-tasklearningstochasticfusiontimeseriescontrastivesmartgrid

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 targets multi-region electric load forecasting (MRELF): predicting short-term electricity demand for many sub-regions at once, where regions share universal load patterns but differ in how much each influencing factor (weather, pricing, holidays) matters. It argues that standard fixes fail for opposite reasons — training one model per region wastes shared signal and overfits, while a single shared model suffers gradient conflicts because different samples demand different emphases on covariates. The central claim is that TriForecaster, which gives experts three distinct specialization axes (one per region, one per context type, one per span of the forecast horizon) and fuses them with a stochastic per-position selection, reduces forecast error by 22.4% on average versus the second-best method on four real-world datasets spanning city grids, zone stations, and buses. If that holds, a single generally applicable model can replace per-region custom builds; the paper reports it already runs in daily production for 17 cities in eastern China serving over 110 million people, outperforming the previous per-city system while saving an estimated 300 thousand U.S. dollars per day.

What carries the argument

At the core is the stochastic fusion module. For each temporal position and latent dimension, the model builds a categorical distribution over experts from pairwise distances between activations (Eqs. 1–2) and draws a hard sample $e \sim \mathrm{Categorical}(\mathbf{P})$ to pick the fused output (Eq. 5). This replaces task-specific gating networks and load-balancing auxiliary losses — the standard MoE machinery — with parameter-free selection driven by batch statistics, letting one framework specialize experts across regions (RegionMixer), contexts (ContextMoE), and time spans (TimeMoE). A contrastive loss (Eq. 10) clusters samples with similar weather, pricing, and date covariates toward th

What would settle it

On the EPC dataset, replace the hard categorical draw in Eq. 5 with a deterministic softmax-weighted average of expert outputs, keeping everything else identical; if MSE does not rise, the discrete routing is not what carries the result. In the same runs, record the marginal distribution of selected experts across positions, contexts, and hours: if the draws stay close to uniform rather than concentrating on interpretable groups (e.g., dawn/midday/evening), the claimed specialization is not happening and the gains come from elsewhere in the architecture.

Watch

Extended reading notes

Core claim

TriForecaster casts each region as a task in a multi-task learning problem. RegionMixer layers run one shared expert plus one expert per region, compute pairwise affinities between expert activations across a batch, and fuse them via a stochastic module that draws a categorical sample per position and latent dimension to pick the winning expert. CTSpecializer layers repeat this pattern on context-specialized and time-specialized experts, with a contrastive loss clustering contextually similar samples. The full system beats all baselines on all four datasets (22.4% average MSE reduction over second-best), and removing any one specialization axis degrades accuracy.

Load-bearing premise

The load-bearing premise is that the stochastic fusion routing is actually trainable: the model picks an expert with a hard categorical draw ($e \sim \mathrm{Categorical}(\mathbf{P})$, Eq. 5), and the paper does not specify how gradients flow through that discrete choice. If it relies on an unstated gradient estimator and that estimator fails in practice, the specialization — and the 22.4% error reduction built on it — would not materialize.

Editorial extensions

If this is right

A single shared architecture can replace per-region model builds across scales: the same TriForecaster configuration topped datasets from 10-minute zone-station loads (EPC) to 15-minute city-level (City-load) and bus-level (Bus-load) loads.
Specialization along all three axes is necessary: the ablation on EPC shows removing RegionMixer, ContextMoE, or TimeMoE each raises error, so the gains are not attributable to any single module.
The approach scales to production: deployed since June 2024 in the eForecaster platform for 17 eastern-China cities (over 110 million people, over 100 GWh daily), it reportedly beats the previous per-city model by 1.6% in accuracy, saving an estimated 2 million kWh and about 300 thousand U.S. dollars per day.
MoE without gating networks transfers from vision and language to time-series regression: TriForecaster outperforms MLoRE, a vision-domain low-rank MoE method, on every dataset, and the authors note MLoRE's low-rank experts are orthogonal and can be integrated into their framework.
The contrastive-loss routing is a controllable design choice: increasing the contrastive coefficient $\alpha$ or the number of context/time experts improves results, per the hyperparameter sweeps.

Reading between the lines

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

The tri-dimensional recipe — group-specific experts, context-specific experts, and horizon-span experts fused by per-position stochastic selection — is generic and could transfer to other grouped time-series tasks such as multi-store retail demand or per-fleet energy loads, wherever groups share dynamics but differ in covariate sensitivity.
The contrast between the 22.4% benchmark MSE reduction and the 1.6% accuracy gain reported in live deployment suggests part of the backtested gain is absorbed by operational realities (data cleaning, postprocessing, domain-knowledge overrides); a natural follow-up is to measure where the gains are lost at each pipeline stage.
Because the fusion draws a hard categorical sample per position and dimension, the model yields an expert-usage map as a byproduct; a testable prediction is that the marginal draws will settle into interpretable patterns — time-mixing experts concentrating on dawn, midday, and evening slices — which grid operators could read as an explanation of each forecast.

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.

Desk Editor's Note

read the letter

Know this first: the paper is a genuine applied contribution, not a repackaging of known tricks. TriForecaster is a clean MoE design for multi-region load forecasting — region-specific experts plus a shared expert, followed by context and time MoE layers, with a contrastive loss. That combination is new for MRELF. The four datasets run from bus-level to city-level, and the eastern China deployment on 17 cities makes the practical claims concrete. The ablations on EPC show each module contributes, and the reported standard deviations are consistently below the baselines', which supports stability.

Where the soft spots actually are, in order. The stress-test's worry about Eq. (5) is the right question but the wrong conclusion. Yes, a hard categorical sample gives no gradient through the sampling to the routing probabilities, and the paper never specifies a straight-through or Gumbel-Softmax estimator. But the router here is not a learned gate — it's a softmax over pairwise activation distances, computed from the experts' own outputs. The selected expert still receives gradients through its output, so the model trains and the routing adapts indirectly as experts move. The real problem is the specification gap: the mechanism as written differs from the SOFTS stochastic pooling they cite, and no inference-time averaging over the sample is stated. So the 22.4% claim rests on an undocumented implementation detail. That is a major-revision issue, not a refutation, and the authors should settle it by stating the estimator and releasing code.

The rest is softer but real. The 22.4% is the mean of per-dataset relative improvements against a second-best baseline that changes per dataset — legitimate aggregation, slightly flattering. The baseline set is small and every baseline uses a TSMixer backbone; the GCN-LSTM adaptation replaces the LSTM, so it is not the published method. The ablation runs on one dataset only. The deployment economics (2 GWh/day, ~$300k/day) are asserted without derivation. The paper has numerous typos and a placeholder ACM template, suggesting a very early preprint. On the citation front, the self-citations to eForecaster are fair — that is the platform the system actually runs on.

Bottom line: this deserves a serious referee. The right outcome is major revision, not desk rejection. For power-systems forecasters and anyone doing MoE in multi-task time series, the architectural idea and the deployed evidence are worth engaging with. If the fusion mechanism gets specified and the code appears, I'd take the 22.4% much more seriously.

Referee Report

4 major / 6 minor

Summary. The paper proposes TriForecaster, a mixture-of-experts framework for multi-region electric load forecasting (MRELF). The model stacks RegionMixer layers, which combine region-specific, region-generic, and pairwise expert information via a stochastic fusion mechanism, followed by CTSpecializer layers that address contextual and temporal variation through context- and time-specializing MoE modules. The authors evaluate on four real-world datasets (EPC, CEESC, City-load, Bus-load) and report that TriForecaster achieves an average 22.4% MSE reduction and 12.3% MAE reduction over the second-best method. They also describe a deployment in eastern China serving 17 cities. The paper's central claim is empirical: the proposed architecture outperforms existing MRELF baselines on the reported datasets.

Significance. If the empirical results are reliable, TriForecaster is a useful contribution to multi-region load forecasting, a problem with practical operational value. The architecture is interesting in that it replaces learned gating networks with a parameter-efficient stochastic fusion based on pairwise expert affinities, which could reduce training complexity and load-balancing issues typical of MoE models. The deployment on the eForecaster platform adds credibility and indicates practical scalability. However, the paper's headline improvement is computed against a per-dataset-varying baseline, and the training/inference behavior of the stochastic fusion is not fully specified; these issues must be clarified before the claims can be fully accepted.

major comments (4)

[§5.2, Table 1] The abstract and Section 5.2 state that TriForecaster achieves an average 22.4% MSE reduction compared to the second-best method. From Table 1, this number is the average of per-dataset relative improvements, where the second-best method changes across datasets: MLoRE for EPC and CEESC, GCN-LSTM for City-load, and MTL for Bus-load. The improvement on City-load is only 7.0% over GCN-LSTM, much smaller than the headline 22.4%. This is not necessarily incorrect, but it is misleading without per-dataset disclosure. Please report per-dataset improvements and clearly state that the average is over dataset-specific second-best baselines. This directly affects the paper's central claim.
[§4.1, Eqs. (1)–(5); §4.2, Eqs. (7)–(9)] The stochastic fusion mechanism draws hard categorical samples, e ~ Categorical(P), and uses the sampled expert output as Opair. The paper does not specify how gradients are handled through this discrete sampling during training, nor how inference is performed (single sample vs. Monte Carlo averaging). Gradients can flow to the parameters of the selected expert through the indexed output, but the sampling operation blocks gradients to P_t. Since P_t is a non-parametric function of expert outputs, this does not strictly prevent training, but it means the selection probabilities are not directly optimized. If the intended mechanism is a differentiable relaxation, the text must state it; if it is hard sampling, the paper must specify the inference procedure (e.g., whether test predictions are averaged over multiple samples) because otherwise the reported MSE values are not deterministic. Th
[§5.3, Table 2] The ablation study is performed only on the EPC dataset, with no standard deviations or significance tests. The text says 'removing any module significantly degrades forecast performance,' but no statistical evidence is given. At minimum, report standard deviations over the five seeds and, ideally, run the ablation on more than one dataset or provide paired significance tests. This is important because the contribution of each architectural component is a key part of the paper's novelty.
[§5.2, Table 3] Although standard deviations are reported, no statistical significance tests are provided for the main comparison. Given the small number of datasets and the varying baselines, a simple paired test (e.g., Wilcoxon signed-rank across datasets or per-dataset t-test over the five seeds) would strengthen the claim that TriForecaster outperforms baselines. This is not a fatal flaw, but it would increase confidence in the headline result.

minor comments (6)

[§1, §7, Appendix A.2] Typos: 'MRELE' in Section 1 should be 'MRELF'; 'provoides' in Section 7; 'gradualy' in Appendix A.2; 'wildly used' in Section 1. Proofreading is needed.
[§4.3, Eq. (13)] The loss uses l_t_i, but Eq. (10) defines l_t_n; notation should be consistent.
[§6] The deployment claim 'outperforms the previous model ... by a margin of 1.6% in terms of average prediction accuracy' is vague. Specify the metric (MSE/MAE) and how the margin is computed.
[§4.1, Ref. [16]] The stochastic fusion mechanism is attributed to SOFTS [16]. Please verify that the cited work indeed proposes stochastic pooling/categorical sampling; the citation currently appears potentially mismatched.
[§2, Baselines] The paper compares against only four baselines. Given the claim of 'state-of-the-art' performance, adding widely used load-forecasting models (e.g., N-BEATS, Informer, PatchTST) would strengthen the comparison, even if the focus is MRELF.
[Eq. (2)] The notation with square brackets in the subscript of Softmax is hard to parse. Please rewrite for clarity.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the 22.4% improvement is an empirical benchmark result, not a derivation from fitted inputs; self-citations concern deployment/preprocessing and are not load-bearing.

full rationale

The paper's central claim is an empirical architecture comparison: TriForecaster is trained with the MSE plus contrastive objective in Eq. (13) and evaluated on holdout test sets of four external datasets against independent baselines (STL, MTL, MLoRE, GCN-LSTM). None of the model equations (Eqs. 1-12) define a target quantity in terms of itself, and no fitted parameter is renamed as a prediction. The stochastic fusion mechanism in Eq. (5) samples an expert output using probabilities computed from batch expert activations in Eqs. (1)-(2); this is a described mechanism, not a circular reduction. Self-citations appear mainly for the eForecaster platform [42,45] and preprocessing tools [35-37], which support the deployment narrative but are not used to justify the headline accuracy gain. The stochastic-fusion idea is cited to external work SOFTS [16], not to the authors' own prior results. The potential absence of a gradient estimator for the categorical sample in Eq. (5) is a training-completeness/correctness concern, not a circularity, so it does not affect this score. The score of 2 reflects only the presence of minor self-citations that are not load-bearing for the central empirical claim.

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

The central empirical claim depends on several hyperparameters tuned on the validation set and on domain-specific modeling assumptions that are not derived from first principles. No new physical entities are introduced.

free parameters (5)

alpha (contrastive loss weight)
Coefficient of contrastive loss in Eq. (13); tuned on the validation set (Section 5.1 and Appendix A.2).
num_e_per_moe (number of experts per CTSpecializer layer)
Tuned on validation; sensitivity shown in Figure 4.
moe_blocks (number of CTSpecializer layers)
Tuned on validation; sensitivity shown in Figure 4.
N_r (number of RegionMixer layers)
Not reported explicitly; an architecture hyperparameter that affects the empirical result.
d_r (latent dimension) = 16
Set to 16 for all benchmarks (Appendix A.1.2).

assumptions (4)

domain assumption Electric load patterns across cities in a region have pronounced similarity, enabling beneficial sharing of information in a multi-task model.
Motivation in Section 1 and Figure 1; if false, the shared-expert design would not help.
ad hoc to paper The pairwise expert affinity computed from per-sample absolute differences across a mini-batch (Eq. 1) is a stable and informative proxy for regional similarity.
This design choice is introduced in Section 4.1 without theoretical or empirical justification beyond the reported results.
ad hoc to paper A hard categorical sample in the stochastic fusion (Eq. 5) can be trained with standard backpropagation despite the discrete sampling.
The paper does not specify a gradient estimator; this is an unstated training assumption.
domain assumption The last month of data is a representative test period for each dataset.
Experimental protocol in Section 5.1; seasonal or anomalous periods would distort the reported error reductions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TriForecaster: A Mixture of Experts Framework for Multi-Region Electric Load Forecasting with Tri-dimensional Specialization." pith.science (2026). https://pith.science/paper/6IWVBG5Y

@misc{pith2026250809753,
  author       = {Pith},
  title        = {Pith review of: TriForecaster: A Mixture of Experts Framework for Multi-Region Electric Load Forecasting with Tri-dimensional Specialization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6IWVBG5Y}},
  note         = {Machine review of arXiv:2508.09753}
}

read the original abstract

Electric load forecasting is pivotal for power system operation, planning and decision-making. The rise of smart grids and meters has provided more detailed and high-quality load data at multiple levels of granularity, from home to bus and cities. Motivated by similar patterns of loads across different cities in a province in eastern China, in this paper we focus on the Multi-Region Electric Load Forecasting (MRELF) problem, targeting accurate short-term load forecasting for multiple sub-regions within a large region. We identify three challenges for MRELF, including regional variation, contextual variation, and temporal variation. To address them, we propose TriForecaster, a new framework leveraging the Mixture of Experts (MoE) approach within a Multi-Task Learning (MTL) paradigm to overcome these challenges. TriForecaster features RegionMixer and Context-Time Specializer (CTSpecializer) layers, enabling dynamic cooperation and specialization of expert models across regional, contextual, and temporal dimensions. Based on evaluation on four real-world MRELF datasets with varied granularity, TriForecaster outperforms state-of-the-art models by achieving an average forecast error reduction of 22.4\%, thereby demonstrating its flexibility and broad applicability. In particular, the deployment of TriForecaster on the eForecaster platform in eastern China exemplifies its practical utility, effectively providing city-level, short-term load forecasts for 17 cities, supporting a population exceeding 110 million and daily electricity usage over 100 gigawatt-hours.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 39 canonical work pages

[1]

Mobarak Abumohsen, Amani Yousef Owda, and Majdi Owda. 2023. Electrical load forecasting using LSTM, GRU, and RNN algorithms. Energies 16, 5 (2023), 2283

work page 2023
[2]

Mashael M Asiri, Ghadah Aldehim, Faiz Abdullah Alotaibi, Mrim M Alnfiai, Mohammed Assiri, and Ahmed Mahmud. 2024. Short-term load forecasting in smart grids using hybrid deep learning. IEEE Access 12 (2024), 23504–23513

work page 2024
[3]

Zico Kolter, and Vladlen Koltun

Shaojie Bai, J. Zico Kolter, and Vladlen Koltun. 2018. An Empirical Evalua- tion of Generic Convolutional and Recurrent Networks for Sequence Modeling. arXiv:1803.01271

arXiv 2018
[4]

Changchun Cai, Yuanjia Li, Zhenghua Su, Tianqi Zhu, and Yaoyao He. 2022. Short- term electrical load forecasting based on VMD and GRU-TCN hybrid network. Applied Sciences 12, 13 (2022), 6647

work page 2022
[5]

Houhe Chen, Mingyang Zhu, Xiao Hu, Jiarui Wang, Yong Sun, and Jinduo Yang

work page
[6]

Si-An Chen, Chun-Liang Li, Nate Yoder, Sercan O Arik, and Tomas Pfister. 2023. Tsmixer: An all-mlp architecture for time series forecasting. arXiv preprint arXiv:2303.06053 (2023)

arXiv 2023
[7]

Tianqi Chen and Carlos Guestrin. 2016. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining . 785–794

work page 2016
[8]

Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In Interna- tional conference on machine learning . PMLR, 1597–1607

work page 2020

Show all 49 references

[9]

Weiqi Chen, Wenwei Wang, Bingqing Peng, Qingsong Wen, Tian Zhou, and Liang Sun. 2022. Learning to Rotate: Quaternion Transformer for Complicated Periodical Time Series Forecasting. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 146–156

2022
[10]

Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. 2020. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. Advances in Neural Information Processing Systems 33 (2020), 2039–2050

2020
[11]

Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik G Learned-Miller, and Chuang Gan. 2023. Mod-squad: Designing mixtures of experts as modular multi-task learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 1...

2023
[12]

Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555 (2014)

2014 arXiv
[13]

Shu Fan, Kittipong Methaprayoon, and Wei-Jen Lee. 2010. Multi-region load fore- casting considering alternative meteorological predictions. In IEEE PES General Meeting. 1–7. doi:10.1109/PES.2010.5589869

2010
[14]

Kinza Fida, Usman Abbasi, Muhammad Adnan, Sajid Iqbal, and Salah El- deen Gasim Mohamed. 2024. A comprehensive survey on load forecasting hybrid models: Navigating the Futuristic demand response patterns through experts and intelligent systems. Results in Engineering (2024), 102773

2024
[15]

Jean-Baptiste Fiot and Francesco Dinuzzo. 2016. Electricity demand forecasting by multi-task learning. IEEE Transactions on Smart Grid 9, 2 (2016), 544–551

2016
[16]

Lu Han, Xu-Yang Chen, Han-Jia Ye, and De-Chuan Zhan. 2024. SOFTS: Efficient Multivariate Time Series Forecasting with Series-Core Fusion. arXiv preprint arXiv:2404.14197 (2024)

2024 arXiv
[17]

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition . 770–778

2016
[18]

S Hochreiter. 1997. Long Short-term Memory. Neural Computation MIT-Press (1997)

1997
[19]

Maryam Imani. 2021. Electrical load-temperature CNN for residential load forecasting. Energy 227 (2021), 120480

2021
[20]

Lianjie Jiang, Xinli Wang, Wei Li, Lei Wang, Xiaohong Yin, and Lei Jia. 2021. Hybrid multitask multi-information fusion deep learning for household short- term load forecasting. IEEE Transactions on Smart Grid 12, 6 (2021), 5362–5372

2021
[21]

Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. 2017. LightGBM: A Highly Efficient Gradient Boosting Decision Tree. In Advances in Neural Information Processing Systems , I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Ferg...

2017
[22]

Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980 (2014)

2014 arXiv
[23]

Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016)

2016 arXiv
[24]

Corentin Kuster, Yacine Rezgui, and Monjur Mourshed. 2017. Electrical load forecasting models: A critical systematic review. Sustainable Cities and Society 35 (2017), 257–270. doi:10.1016/j.scs.2017.08.009 TriForecaster: A Mixture of Experts Framework for Multi-Region Electric...

2017 doi
[25]

Bo Liu, Xingchao Liu, Xiaojie Jin, Peter Stone, and Qiang Liu. 2021. Conflict- averse gradient descent for multi-task learning. Advances in Neural Information Processing Systems 34 (2021), 18878–18890

2021
[26]

Haizhou Liu, Xuan Zhang, Hongbin Sun, and Mohammad Shahidehpour. 2023. Boosted multi-task learning for inter-district collaborative load forecasting. IEEE Transactions on Smart Grid 15, 1 (2023), 973–986

2023
[27]

CHEN Lüpeng, YIN Linfei, and et al. YU Tao. 2018. Short-term Power Load Forecasting Based on Deep Forest Algorithm[J]. Electric Power Construction 39, 11 (2018), 42–50. doi:10.3969/j.issn.1000-7229.2018.11.006

2018
[28]

Alexandra L’Heureux, Katarina Grolinger, and Miriam AM Capretz. 2022. Transformer-based model for electrical load forecasting. Energies 15, 14 (2022), 4993

2022
[29]

Haris Mansoor, Muhammad Shuzub Gull, Huzaifa Rauf, Muhammad Khalid, Naveed Arshad, et al. 2024. Graph Convolutional Networks based short-term load forecasting: Leveraging spatial information for improved accuracy. Electric Power Systems Research 230 (2024), 110263

2024
[30]

Nada Mounir, Hamid Ouadi, and Ismael Jrhilifa. 2023. Short-term electric load forecasting using an EMD-BI-LSTM approach for smart grid energy management system. Energy and Buildings 288 (2023), 113022

2023
[31]

Boris N Oreshkin, Grzegorz Dudek, Paweł Pełka, and Ekaterina Turkina. 2021. N- BEATS neural network for mid-term electricity load forecasting. Applied Energy 293 (2021), 116918

2021
[32]

Salam and A

A. Salam and A. E. Hibaoui. 2018. Comparison of Machine Learning Algorithms for the Power Consumption Prediction: Case Study of Tetouan city. 2018 6th International Renewable and Sustainable Energy Conference (IRSEC) (2018). doi:10. 1109/irsec.2018.8703007

2018
[33]

Robert H Shumway, David S Stoffer, Robert H Shumway, and David S Stoffer
[34]

Sean J Taylor and Benjamin Letham. 2018. Forecasting at scale. The American Statistician 72, 1 (2018), 37–45

2018
[35]

Qingsong Wen, Jingkun Gao, Xiaomin Song, Liang Sun, Huan Xu, and Shenghuo Zhu. 2019. RobustSTL: A robust seasonal-trend decomposition algorithm for long time series. In AAAI. 1501–1509

2019
[36]

Qingsong Wen, Kai He, Liang Sun, Yingying Zhang, Min Ke, and Huan Xu
[37]

Qingsong Wen, Zhe Zhang, Yan Li, and Liang Sun. 2020. Fast RobustSTL: Effi- cient and Robust Seasonal-Trend Decomposition for Time Series with Complex Patterns. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD) . 2203–2213

2020
[38]

Kailang Wu, Jie Gu, Lu Meng, Honglin Wen, and Jinghuan Ma. 2022. An explain- able framework for load forecasting of a regional integrated energy system based on coupled features and multi-task learning. Protection and Control of Modern Power Systems 7, 2 (2022), 1–14. doi:10.1...

2022 doi
[39]

Yuqi Yang, Peng-Tao Jiang, Qibin Hou, Hao Zhang, Jinwei Chen, and Bo Li. 2024. Multi-Task Dense Prediction via Mixture of Low-Rank Experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 27927– 27937

2024
[40]

Yuqi Yang, Peng-Tao Jiang, Qibin Hou, Hao Zhang, Jinwei Chen, and Bo Li. 2024. Multi-Task Dense Prediction via Mixture of Low-Rank Experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . 27927–27937

2024
[41]

Hanrong Ye and Dan Xu. 2023. Taskexpert: Dynamically assembling multi-task representations with memorial mixture-of-experts. InProceedings of the IEEE/CVF International Conference on Computer Vision . 21828–21837

2023
[42]

Zhu Zhaoyang, Chen Weiqi, Xia Rui, Zhou Tian, Niu Peisong, Peng Bingqing, Wang Wenwei, Liu Hengbo, Ma Ziqing, Wen Qingsong, and Sun Liang. 2023. eForecaster: Unifying Electricity Forecasting with Robust, Flexible, and Explain- able Machine Learning Algorithms. In Thirty-Sevent...

2023
[43]

Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. 2021. Informer: Beyond efficient transformer for long se- quence time-series forecasting. In Proceedings of the AAAI conference on artificial intelligence, Vol. 35. 11106–11115

2021
[44]

Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin
[45]

Zhaoyang Zhu, Weiqi Chen, Rui Xia, Tian Zhou, Peisong Niu, Bingqing Peng, Wenwei Wang, Hengbo Liu, Ziqing Ma, Qingsong Wen, and Liang Sun. 2023. eForecaster: unifying electricity forecasting with robust, flexible, and explainable machine learning algorithms. In Proceedings of ...

2023
[2017]

Time series analysis and its applications: with R examples (2017), 75–163

ARIMA models. Time series analysis and its applications: with R examples (2017), 75–163

2017
[2021]

In Proceedings of the 2021 International Conference on Management of Data(SIGMOD)

RobustPeriod: Robust Time-Frequency Mining for Multiple Periodicity Detection. In Proceedings of the 2021 International Conference on Management of Data(SIGMOD). 2328–2337

2021
[2022]

In International Conference on Machine Learning

Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting. In International Conference on Machine Learning. PMLR, 27268– 27286
[2023]

Energy Reports 9 (2023), 1022–1031

Research on short-term load forecasting of new-type power system based on GCN-LSTM considering multiple influencing factors. Energy Reports 9 (2023), 1022–1031

2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Mobarak Abumohsen, Amani Yousef Owda, and Majdi Owda. 2023. Electrical load forecasting using LSTM, GRU, and RNN algorithms. Energies 16, 5 (2023), 2283

work page 2023

[2] [2]

Mashael M Asiri, Ghadah Aldehim, Faiz Abdullah Alotaibi, Mrim M Alnfiai, Mohammed Assiri, and Ahmed Mahmud. 2024. Short-term load forecasting in smart grids using hybrid deep learning. IEEE Access 12 (2024), 23504–23513

work page 2024

[3] [3]

Zico Kolter, and Vladlen Koltun

Shaojie Bai, J. Zico Kolter, and Vladlen Koltun. 2018. An Empirical Evalua- tion of Generic Convolutional and Recurrent Networks for Sequence Modeling. arXiv:1803.01271

arXiv 2018

[4] [4]

Changchun Cai, Yuanjia Li, Zhenghua Su, Tianqi Zhu, and Yaoyao He. 2022. Short- term electrical load forecasting based on VMD and GRU-TCN hybrid network. Applied Sciences 12, 13 (2022), 6647

work page 2022

[5] [5]

Houhe Chen, Mingyang Zhu, Xiao Hu, Jiarui Wang, Yong Sun, and Jinduo Yang

work page

[6] [6]

Si-An Chen, Chun-Liang Li, Nate Yoder, Sercan O Arik, and Tomas Pfister. 2023. Tsmixer: An all-mlp architecture for time series forecasting. arXiv preprint arXiv:2303.06053 (2023)

arXiv 2023

[7] [7]

Tianqi Chen and Carlos Guestrin. 2016. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining . 785–794

work page 2016

[8] [8]

Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In Interna- tional conference on machine learning . PMLR, 1597–1607

work page 2020