Pith. sign in

REVIEW 4 major objections 4 minor 31 references

A pattern-driven method estimates each grid entity's demand-response flexibility from its own historical load curves and cuts regulation deviation by up to 91.87%.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 11:45 UTC pith:NB75VJ3T

load-bearing objection The clustering-plus-MILP idea is worth a look, but the evaluation never defines 'real load reduction,' and that gap guts the headline performance claims. the 4 major comments →

arxiv 2607.22590 v1 pith:NB75VJ3T submitted 2026-06-10 math.OC cs.LG

DRP-FLR: Data-Driven Assessment of Demand Response Potential for Flexible Load Regulation in Smart Grids

Yunhao Yao , Siyu Jing , Yang Yang , Qiang Xu , Changqi Weng , Xiang-Yang Li This is my paper
classification math.OC cs.LG MSC 90C1190C10
keywords demand responseflexible load regulationload forecastingconsumption pattern clusteringDBSCANmixed-integer linear programmingsmart gridregulation capacity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper claims that the flexible load an entity can shed in a demand-response event equals the gap between the mean and the minimum historical load within the entity's matched consumption-pattern cluster at the target hour. Embedding that estimate into a mixed-integer selector of participants, and weighting renewable-friendly entities, yields a regulation plan that meets grid demand without over- or under-regulating. On a regional grid and a campus microgrid, the authors report deviation reduced by 36.63%-91.87% compared with baselines, and participant benefit up by 44.66% on average. If true, the method makes accurate demand-response potential available from passive load measurements alone, with no need for before/after response labels.

Core claim

The central claim is that consumption-pattern clustering, not pre/post DR labels, carries the information needed to quantify an entity's load regulation capacity. For each entity, historical daily load curves are clustered with a density-based algorithm; the forecasted day-ahead profile is assigned to a cluster; and the regulation capacity at a given hour is defined as the mean minus the minimum load over that cluster's historical days at the same hour. The demand-response potential is the ratio of this capacity to the contracted response amount. A 0-1 mixed-integer linear program then selects which entities to call, maximizing aggregated potential and participant benefit under a bounded reg

What carries the argument

The central object is the pattern-clustered mean-minus-min regulation capacity R_k(M+δ_T) = μ_cluster(δ_T) − ℓ_cluster(δ_T), where μ and ℓ are the mean and minimum load within the entity's matched consumption-pattern cluster at the regulation hour. This identity converts a cluster of historical load curves into a per-hour estimate of how much the entity can be expected to shed, without ever observing an actual demand-response event. It feeds the DR potential ratio P_k = R_k / R_c, which in turn drives the MILP selector. The DBSCAN clustering with silhouette-based hyperparameter selection is the mechanism that identifies which historical days are comparable to the forecasted day.

Load-bearing premise

The evaluation assumes that the 'real load reductions' R_real_k used to compute the Achievement Rate are independent post-event measurements, while the datasets are described as passive consumption/generation measurements with no demand-response event labels; if R_real_k is instead derived from the same estimated mean-minus-min capacity, the reported Achievement Rate is forced toward the target window.

What would settle it

Run DRP-FLR on a dataset that contains actual demand-response events with measured pre- and post-event consumption for the same entities, and compare the predicted mean-minus-min capacity against the realized load reduction event by event. If the correlation is no better than a global historical average, or if the Achievement Rate computed from truly independent measurements leaves the [1, 1+σ] window as often as the baselines, the central claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Grid operators can estimate entity-level flexibility from passive smart-meter data alone, removing the need for labeled demand-response events.
  • Entity-specific pattern profiles correct both systematic over-estimation (long-tail submeter loads) and under-estimation (uniform enterprise loads) that global-average baselines suffer.
  • The MILP formulation with soft shortfall/slack penalties guarantees a feasible regulation plan even when flexibility is scarce or forecasts err.
  • Weighting renewable-friendly entities is an explicit lever to increase renewable accommodation within the same regulation optimization.
  • Reported deviations of [-4.95%, +13.53%] on the regional grid and [-5.49%, +2.71%] on the microgrid sit close to the ideal [100%, 120%] achievement window.

Where Pith is reading between the lines

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

  • The mean-minus-min identity implies that an entity's minimum historical load at an hour encodes its non-sheddable floor; if that floor is robust, the same cluster statistics could be converted into quantile-based confidence intervals for regulation capacity, giving operators risk-bounded reserves.
  • Because clustering is per-entity, the pipeline transfers to a new grid using only historical load curves; a natural test is to benchmark it on a third-party dataset with actual demand-response events to independently verify the 36-92% deviation reduction.
  • The soft-constraint MILP could be extended to a rolling multi-period horizon with online cluster updates, which might smooth the Achievement Rate overshoot observed at small regulation ratios (e.g., 118-134% at 6% demand on the regional grid).
  • The weighting lever suggests a direct extension to jointly optimizing renewable curtailment and DR dispatch, where the renewable weight is learned rather than set a priori.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes DRP-FLR, a data-driven pipeline for demand response (DR) potential assessment and flexible load regulation. It forecasts load with TimeGPT enriched by exogenous embeddings, clusters historical daily load curves with DBSCAN to identify entity-specific consumption patterns, estimates each entity's regulation capacity R_k as the mean-minus-minimum load within the matched pattern cluster, and then solves a mixed-integer linear program (MILP) to select participants, jointly maximizing DR potential, participant benefit, and renewable-weighted participation while enforcing a bounded total regulation. Experiments on a regional grid dataset (441 enterprises) and a campus microgrid dataset (73 submeters) report Achievement Rates mostly inside the target interval and claim a 36.63%–91.87% reduction in regulation deviation and a 44.66% average improvement in participant benefit relative to baselines.

Significance. If the central empirical claim were supported, the paper would offer a practically attractive idea: replacing hard-to-obtain pre/post DR labels with pattern-clustered statistical ranges as DR potential estimates, embedded in a clean and solvable MILP. The formulation itself is clear, and the use of two real-world datasets is a strength. However, the evaluation does not currently establish the claim. The metric that drives all headline numbers, the 'real load reduction' R_real_k, is never defined or sourced, and the datasets are described only as passive measurements with no DR event labels. The paper also reports no forecast-accuracy validation, no uncertainty quantification, and no code/data release. Because the empirical contribution depends on an unmeasured ground truth, the significance of the reported 36.63%–91.87% and 44.66% numbers cannot be assessed. The underlying idea and MILP model have potential, but the present evidence is insufficient for publication.

major comments (4)
  1. [Section V.A.2 (Evaluation Metrics)] The Achievement Rate is defined as AR = (Σ_k S_k R_real_k)/L, where R_real_k is called the 'real load reduction delivered by entity k', but the paper never specifies how R_real_k is obtained. The datasets in Section V.A.1 are described only as 15-minute load/generation measurements with no demand-response event labels or counterfactual baselines. This is load-bearing: if R_real_k is set equal to the estimated R_k = μ − min used in the MILP, then the constraint L ≤ Σ S_k R_k ≤ (1+σ)L forces AR into [1,1+σ] up to soft-penalty violations, and the reported improvement over baselines reflects constraint satisfaction rather than accurate DR potential estimation. If R_real_k is an independent post-intervention measurement, the measurement protocol, event definition, and baseline calculation must be described. Without this, the central empirical claim is ungrounded.
  2. [Section IV.B.2 / Definition IV-B.1] The regulation capacity R_k(M+δ_T) = μ_cluster(δ_T) − ℓ_cluster(δ_T) is asserted to represent the entity's feasible load-reduction range. This is a strong construct-validity assumption: there is no evidence in the paper that the historical mean-minus-minimum within a pattern cluster corresponds to what an entity can actually deliver in a DR event. The same quantity is used to define P_k, B_k, the MILP objective, and the regulation constraints. No comparison with measured reductions, no sensitivity analysis on the capacity definition, and no alternative estimator is reported. Since the evaluation metric R_real is not independently sourced, the paper cannot validate this central assumption.
  3. [Section V.B (Tables I–IV)] All reported results are single-point values with no error bars, confidence intervals, or statistical tests. No number of evaluation days/events, no temporal train/test split, and no forecast-accuracy metrics are provided, even though the whole pipeline depends on forecasted loads. For example, Table III's regional-grid comparison at a 10% ratio (95.22% vs. 125.17% Achievement Rate) is presented as deterministic, and the claimed 36.63%–91.87% reduction range appears to be derived from a small set of comparisons. The reader cannot tell whether these numbers are robust or reflect a few favorable days. The absence of uncertainty quantification is especially problematic for a data-driven method whose performance depends on forecast errors and stochastic load behavior.
  4. [Section V.A.1 / Definition IV-B.1 / Definition IV-C.1] The definitions of DR potential P_k and benefit B_k require the contracted response amount R_c_k, but the paper does not state how R_c_k is obtained for the datasets. The datasets are described as passive measurements; no contract information is mentioned. Since the MILP objective uses P_k and B_k, and since B_k uses thresholds relative to R_c_k (0.8R_c, 1.2R_c), the experimental setup is incomplete. If R_c_k is assumed, estimated, or set synthetically, this must be stated explicitly; otherwise the objective function cannot be computed on the described data.
minor comments (4)
  1. [Tables III and IV] Tables III and IV contain the typo 'Ragulation Ratio' instead of 'Regulation Ratio'.
  2. [Section V.A.1] The text uses 'HV AC' where 'HVAC' is intended, and the sentence at the top of page 10 should read 'are reward-effective' rather than 'are reward-effective' (minor wording issue).
  3. [Section IV.A] The notation ']Emb' appears in the forecasting pipeline equation; this appears to be a typo for the fused embedding variable.
  4. [Section V.B / IV.C.3] The paper never reports the exact values of the penalty coefficients λ_ξ and λ_ζ, nor the default value of α across all experiments. These parameters affect the degree to which the soft constraints are violated and are needed for reproducibility.

Circularity Check

2 steps flagged

Evaluation metrics AR and PB are built from the same estimated capacity R_k that the MILP optimizes; the headline performance figures are not independently grounded.

specific steps
  1. fitted input called prediction [Section V.A.2 (Evaluation Metrics) and Definition IV-C.1]
    "Let B_real_k denote the real load regulation of entity k, computed according to Definition IV-C.1. Then Participant Benefit is defined as: PB = Σ S_k · B_real_k."

    Definition IV-C.1 defines B_k(t) exclusively as a piecewise function of R_k(t), the estimated regulation capacity from Section IV.B.2. Thus 'B_real' computed 'according to Definition IV-C.1' is the same estimated benefit that already appears in the MILP objective term Σ S_k B_k W_k. The reported '44.66% average participant benefit improvement' is therefore a comparison of the optimizer's own internal benefit proxy, not an independently measured monetary reward. The evaluation metric reduces to the maximized objective by construction.

  2. fitted input called prediction [Section V.A.2, with Section IV.B.2 and Section IV.C.3]
    "Denote the real load reduction delivered by entity k as R_real_k ≥ 0. Then the Achievement Rate is defined as: AR = ... We replace the hard bounds with the following soft constraints: sum_R + ξ ≥ L, sum_R ≤ (1+σ)L + ζ."

    The paper never defines or sources R_real_k. Section V.A.1 describes only passive 15-minute load/generation measurements, with no demand-response event labels or counterfactual baselines. The only load-reduction quantity defined anywhere is R_k = μ_cluster − min_cluster from the pattern-cluster estimation. Substituting that defined quantity gives AR = Σ S_k R_k / L, which the soft constraints with large penalties λ_ξ, λ_ζ force toward [1, 1+σ]. The AR values in Tables I and II then certify constraint satisfaction rather than independent DR-potential accuracy. The paper's statement that 'forecasting errors can lead to discrepancies between predicted and real load reductions' asserts independence, but no independent real delivery is measured or defined.

full rationale

The paper's forecast and clustering modules are externally based (TimeGPT, DBSCAN, silhouette) and do not by themselves create circularity. The circularity arises in the evaluation layer: both headline metrics—Achievement Rate and Participant Benefit—are defined using the same estimated reduction R_k that the MILP selects on. The benefit metric explicitly reuses Definition IV-C.1, which is a function of the estimated R_k, so PB is essentially the optimized objective term relabeled as 'real'. The achievement-rate metric uses an undefined R_real_k; with no DR-event labels or post-intervention measurements in the described datasets, the only available reduction quantity is the estimated R_k. If R_real_k equals R_k, AR is constrained by the MILP soft constraints to the target interval, making the reported 36.63%–91.87% deviation reduction a consequence of the constraint, not of predictive accuracy. This affects the paper's central empirical claims. I do not see a load-bearing self-citation chain; the problem is self-definitional evaluation rather than citation circularity. Score 7 reflects that the core empirical validation reduces to the model's own quantities, though an external R_real source, if described, could rescue the AR claim.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The core estimator R_k = mu - min is an inductive bias with no physical validation, and the evaluation additionally assumes access to 'real' reductions that are never sourced. The free parameters alpha, sigma, lambda, and the per-entity DBSCAN selection all shape the reported numbers, making the empirical claims depend on an unstated data-generation process.

free parameters (4)
  • trade-off factor alpha = varied 1e1 to 1e4 in experiments
    Balances DR potential versus participant benefit in the scalarized MILP objective; no principled choice is given, and results shift materially with alpha.
  • over-regulation tolerance sigma = 0.2
    Sets the acceptable upper bound [1, 1+sigma] on Achievement Rate; chosen by hand and not varied in sensitivity analysis.
  • soft-constraint penalties lambda_xi, lambda_zeta = not specified (stated as >> 1)
    Penalize shortfall and overshoot to keep the MILP feasible; exact magnitudes are never reported, so the trade-off between feasibility and target adherence is uncalibrated.
  • DBSCAN hyperparameters (epsilon, minPts) = epsilon from 25/50/75% quantiles of mean 4-NN distance; minPts in {3,4}
    Selected per entity by maximizing silhouette on the same historical data, effectively fitting the clustering configuration to the data used for evaluation.
axioms (4)
  • domain assumption Historical days assigned to the same DBSCAN cluster share one load pattern, and the cluster's mean-minus-min at hour delta is a valid estimate of that entity's feasible load-reduction capacity at a future hour.
    Section IV.B.2 defines R_k = mu - min; there is no physical or empirical validation that an entity can shed load down to the historical minimum, nor that future behavior follows the matched cluster.
  • domain assumption The datasets contain or can yield ground-truth 'real load reductions' R_real_k used to compute Achievement Rate and Participant Benefit.
    Section V.A.2 introduces R_real_k but never says where it comes from; dataset descriptions in Section V.A.1 mention only passive consumption/generation measurements, with no demand-response event labels.
  • domain assumption A forecasted load profile can be assigned to the nearest non-noise historical cluster, and that mapping remains valid across the forecast horizon.
    Section IV.B.2 maps the predicted profile PL_k to the nearest cluster; forecast errors can misassign patterns and invalidate the mean-min range.
  • domain assumption The pretrained TimeGPT backbone behaves as described and its embeddings, combined with the exogenous-knowledge embeddings, produce accurate hourly forecasts.
    Section IV.A relies on TimeGPT [28] as the forecasting backbone, but the paper reports no independent forecast-accuracy evaluation or ablation of the exogenous embedding.

pith-pipeline@v1.3.0-alltime-deepseek · 14350 in / 12153 out tokens · 132647 ms · 2026-08-02T11:45:50.142119+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of DRP-FLR: Data-Driven Assessment of Demand Response Potential for Flexible Load Regulation in Smart Grids." pith.science (2026). https://pith.science/paper/NB75VJ3T

@misc{pith2026260722590,
  author       = {Pith},
  title        = {Pith review of: DRP-FLR: Data-Driven Assessment of Demand Response Potential for Flexible Load Regulation in Smart Grids},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NB75VJ3T}},
  note         = {Machine review of arXiv:2607.22590}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The rapid growth of AI workloads and renewable energy resources exacerbates supply-demand imbalance in power systems, making traditional load regulation designed for efficient allocation inadequate and motivating demand response (DR) mechanisms to enable load controllability in smart grids. However, existing DR-oriented approaches either focus on optimizing electricity cost or occupant comfort with limited benefit to system-level balance. Others overlook the diverse and dynamic consumption patterns of heterogeneous energy entities, leading to significant over- or under-regulation. Therefore, we propose DRP-FLR. First, DRP-FLR achieves accurate short-term load forecasting by embedding exogenous knowledge (e.g., entity information, prediction time) into historical load representations. Next, it constructs entity-specific load-pattern profiles by clustering historical load curves, and estimates DR potential by matching forecasted loads with pattern profiles. Finally, DRP-FLR formulates flexible load regulation as a mixed-integer optimization problem and solves it with an MILP solver to jointly optimize DR utilization, participant economic benefit, and renewable accommodation, while enforcing supply-demand balance and economic feasibility. Experiments on a regional grid and a campus microgrid show that DRP-FLR reduces regulation deviation by 36.63%-91.87% and improves participant benefit by 44.66% on average.

Figures

Figures reproduced from arXiv: 2607.22590 by Changqi Weng, Qiang Xu, Siyu Jing, Xiang-Yang Li, Yang Yang, Yunhao Yao.

Figure 1
Figure 1. Figure 1: The Process of Demand Response Mechanism [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The System Overview of DRP-FLR type, forecasting time period, and workday/holiday indicator into learnable embeddings, and overlay them onto the embed￾dings of the historical load sequence. The fused embeddings are fed into the backbone Fpre(·) for hourly load prediction. Let Entk , Time, and DayType denote the exogenous knowledge, and let Embk ent, Embtime and EmbdayType be their embeddings. Let Loadk [M … view at source ↗
Figure 3
Figure 3. Figure 3: Example of Demand Response Potential: An Estate Company in A Regional Grid [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Example of Demand Response Potential: An Office Area in A Microgrid [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison with Existing Demand Response Potential Estimation Methods. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison with Existing Naive Load Regulation Strategies. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

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

31 extracted references · 2 linked inside Pith

  1. [1]

    Overview of demand-response services: A review,

    D. Stanelyte, N. Radziukyniene, and V . Radziukynas, “Overview of demand-response services: A review,”Energies, vol. 15, no. 5, p. 1659, 2022

    2022

  2. [2]

    Impacts of digitalization on smart grids, renewable energy, and demand response: An updated review of current applica- tions,

    M. Mahmood, P. Chowdhury, R. Yeassin, M. Hasan, T. Ahmad, and N.- U.-R. Chowdhury, “Impacts of digitalization on smart grids, renewable energy, and demand response: An updated review of current applica- tions,”Energy Conversion and Management: X, vol. 24, p. 100790, 2024

    2024

  3. [3]

    Ai-empowered methods for smart energy consumption: A review of load forecasting, anomaly detection and demand response,

    X. Wang, H. Wang, B. Bhandari, and L. Cheng, “Ai-empowered methods for smart energy consumption: A review of load forecasting, anomaly detection and demand response,”International Journal of Precision Engineering and Manufacturing-Green Technology, vol. 11, no. 3, pp. 963–993, 2024

    2024

  4. [4]

    Multi-agent deep reinforcement learning based demand response and energy management for heavy industries with discrete manufacturing systems,

    A. Bashyal, T. Boroukhian, P. Veerachanchai, M. Naransukh, and H. Wicaksono, “Multi-agent deep reinforcement learning based demand response and energy management for heavy industries with discrete manufacturing systems,”Applied Energy, vol. 392, p. 125990, 2025

    2025

  5. [5]

    Minlp probabilistic scheduling model for demand response programs integrated energy hubs,

    M. Alipour, K. Zare, and M. Abapour, “Minlp probabilistic scheduling model for demand response programs integrated energy hubs,”IEEE Transactions on Industrial Informatics, vol. 14, no. 1, pp. 79–88, 2017

    2017

  6. [6]

    Residential demand response scheduling with multiclass appliances in the smart grid,

    H.-T. Roh and J.-W. Lee, “Residential demand response scheduling with multiclass appliances in the smart grid,”IEEE Transactions on Smart Grid, vol. 7, no. 1, pp. 94–104, 2015

    2015

  7. [7]

    Eshraghi, G

    A. Eshraghi, G. Salehi, S. Heibati, and K. Lari, “An enhanced operation model for energy storage system of a typical combined cool, heat and power based on demand response program: The application of mixed integer linear programming,”Building Services Engineering Research and Technology, vol. 40, no. 1, pp. 47–74, 2019

    2019

  8. [8]

    A multi-objective demand response optimization model for scheduling loads in a home energy management system,

    J. M. Veras, I. R. S. Silva, P. R. Pinheiro, R. A. Rab ˆelo, A. F. S. Veloso, F. A. S. Borges, and J. J. Rodrigues, “A multi-objective demand response optimization model for scheduling loads in a home energy management system,”Sensors, vol. 18, no. 10, p. 3207, 2018

    2018

  9. [9]

    Online transfer learning-based residential demand response potential forecasting for load aggregator,

    K. Li, Z. Li, C. Huang, and Q. Ai, “Online transfer learning-based residential demand response potential forecasting for load aggregator,” Applied Energy, vol. 358, p. 122631, 2024

    2024

  10. [10]

    Integration of demand response and short-term forecasting for the management of prosumers’ demand and generation,

    M. C. Ruiz-Abell ´on, L. A. Fern´andez-Jim´enez, A. Guillam´on, A. Falces, A. Garc´ıa-Garre, and A. Gabald ´on, “Integration of demand response and short-term forecasting for the management of prosumers’ demand and generation,”Energies, vol. 13, no. 1, p. 11, 2019

    2019

  11. [11]

    Day- ahead demand response potential prediction in residential buildings with hitskan: A fusion of kolmogorov-arnold networks and n-hits,

    A. Muqtadir, B. Li, Z. Ying, C. Songsong, and S. N. Kazmi, “Day- ahead demand response potential prediction in residential buildings with hitskan: A fusion of kolmogorov-arnold networks and n-hits,”Energy and Buildings, vol. 332, p. 115455, 2025

    2025

  12. [12]

    Applications of probabilistic forecasting in demand response,

    M. C. Ruiz-Abell ´on, L. A. Fern ´andez-Jim´enez, A. Guillam ´on, and A. Gabald ´on, “Applications of probabilistic forecasting in demand response,”Applied Sciences, vol. 14, no. 21, p. 9716, 2024

    2024

  13. [13]

    Comprehensive forecasting of california’s energy consumption: A multi-source and sectoral analysis using arima and arimax models,

    Z. Moslemi, L. Clark, S. Kernal, S. Rehome, S. Sprengel, A. Tamizifar, S. Tuli, V . Chokshi, M. Nomeli, E. Lianget al., “Comprehensive forecasting of california’s energy consumption: A multi-source and sectoral analysis using arima and arimax models,”arXiv preprint arXiv:2402.04432, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Time-series clustering and forecasting household electricity demand using smart meter data,

    H. Kim, S. Park, and S. Kim, “Time-series clustering and forecasting household electricity demand using smart meter data,”Energy Reports, vol. 9, pp. 4111–4121, 2023

    2023

  15. [15]

    Nonlinear arimax model for long–term sectoral demand forecasting,

    N. Neshat, H. Hadian, and M. Behzad, “Nonlinear arimax model for long–term sectoral demand forecasting,”Management Science Letters, vol. 8, no. 6, pp. 581–592, 2018

    2018

  16. [16]

    Short-term load forecasting based on deep learning for end-user transformer subject to volatile electric heating loads,

    Q. Chen, M. Xia, T. Lu, X. Jiang, W. Liu, and Q. Sun, “Short-term load forecasting based on deep learning for end-user transformer subject to volatile electric heating loads,”IEEE Access, vol. 7, pp. 162 697– 162 707, 2019

    2019

  17. [17]

    An approach for demand forecasting in steel industries using ensemble learning,

    S. T. U. Raju, A. Sarker, A. Das, M. M. Islam, M. S. Al-Rakhami, A. M. Al-Amri, T. Mohiuddin, and F. R. Albogamy, “An approach for demand forecasting in steel industries using ensemble learning,”Complexity, vol. 2022, no. 1, p. 9928836, 2022

    2022

  18. [18]

    Enhancing short-term probabilistic load forecasting and scenario generation with tailored kernel functions in mixture density networks,

    T. Ochoa, C. Serpell, C. Valle, and E. Gil, “Enhancing short-term probabilistic load forecasting and scenario generation with tailored kernel functions in mixture density networks,”Expert Systems with Applications, vol. 284, p. 127932, 2025

    2025

  19. [19]

    Enhancing peak electricity demand forecasting for commercial build- ings using novel lstm loss functions,

    H. S. Nyg ˚ard, S. Grøtan, K. R. Kvisberg, L. R. Gorjao, and T. Martinsen, “Enhancing peak electricity demand forecasting for commercial build- ings using novel lstm loss functions,”Electric Power Systems Research, vol. 246, p. 111722, 2025

    2025

  20. [20]

    Predicting electricity con- sumption for commercial and residential buildings using deep recurrent neural networks,

    A. Rahman, V . Srikumar, and A. D. Smith, “Predicting electricity con- sumption for commercial and residential buildings using deep recurrent neural networks,”Applied energy, vol. 212, pp. 372–385, 2018

    2018

  21. [21]

    A short-term load forecasting model of lstm neural network considering demand response,

    X. Guo, Q. Zhao, S. Wang, D. Shan, and W. Gong, “A short-term load forecasting model of lstm neural network considering demand response,” Complexity, vol. 2021, no. 1, p. 5571539, 2021

    2021

  22. [22]

    A transformer based approach to electricity load forecasting,

    J. W. Chan and C. K. Yeo, “A transformer based approach to electricity load forecasting,”The Electricity Journal, vol. 37, no. 2, p. 107370, 2024

    2024

  23. [23]

    Short-term load forecasting based on the transformer model,

    Z. Zhao, C. Xia, L. Chi, X. Chang, W. Li, T. Yang, and A. Y . Zomaya, “Short-term load forecasting based on the transformer model,” information, vol. 12, no. 12, p. 516, 2021

    2021

  24. [24]

    Stacked hybrid model for load forecasting: integrating transformers, ann, and fuzzy logic,

    A. Balakrishnan, B. Sanisetty, R. B. Bandaruet al., “Stacked hybrid model for load forecasting: integrating transformers, ann, and fuzzy logic,”Scientific Reports, vol. 15, no. 1, pp. 1–24, 2025

    2025

  25. [25]

    Short-term electricity grid maximum demand forecasting with the arimax-svr machine learning hybrid model,

    H. F. Chow, “Short-term electricity grid maximum demand forecasting with the arimax-svr machine learning hybrid model,”Hkie Trans, vol. 28, pp. 22–30, 2021

    2021

  26. [26]

    A hybrid model based on selective ensemble for energy consumption forecasting in china,

    J. Xiao, Y . Li, L. Xie, D. Liu, and J. Huang, “A hybrid model based on selective ensemble for energy consumption forecasting in china,” Energy, vol. 159, pp. 534–546, 2018

    2018

  27. [27]

    Hybrid ensemble intelligent model based on wavelet transform, swarm intelligence and artificial neural network for electricity demand forecasting,

    E. O.-N. Jnr, Y . Y . Ziggah, and S. Relvas, “Hybrid ensemble intelligent model based on wavelet transform, swarm intelligence and artificial neural network for electricity demand forecasting,”Sustainable Cities and Society, vol. 66, p. 102679, 2021

    2021

  28. [28]

    Timegpt-1,

    A. Garza, C. Challu, and M. Mergenthaler-Canseco, “Timegpt-1,”arXiv preprint arXiv:2310.03589, 2023

    Pith/arXiv arXiv 2023

  29. [29]

    A density-based algorithm for discovering clusters in large spatial databases with noise,

    M. Ester, H.-P. Kriegel, J. Sander, X. Xuet al., “A density-based algorithm for discovering clusters in large spatial databases with noise,” inkdd, vol. 96, no. 34, 1996, pp. 226–231

    1996

  30. [30]

    Silhouettes: a graphical aid to the interpretation and validation of cluster analysis,

    P. J. Rousseeuw, “Silhouettes: a graphical aid to the interpretation and validation of cluster analysis,”Journal of computational and applied mathematics, vol. 20, pp. 53–65, 1987

    1987

  31. [31]

    Notice on issuing the{Power Demand Side Management Measures (2023 Edition)},

    National Development and Reform Commission of China, “Notice on issuing the{Power Demand Side Management Measures (2023 Edition)},” National Development and Reform Commission of China website, Sep. 2023, document No.: Fa Gai Yun Xing Gui [2023] 1283; Published: 2023-09-27. [Online]. Available: https: //www.ndrc.gov.cn/xxgk/zcfb/ghxwj/202309/t20230927 1360902.html

    2023