Pith. sign in

REVIEW 4 major objections 5 minor 50 references

External Data-Enhanced Meta-Representation for Adaptive Probabilistic Load Forecasting

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper argues that external data should act as meta-knowledge that dynamically rewires a load-forecasting model's parameters, and reports that the resulting M2oE2 framework cuts test error to roughly 5-50% of standard baselines.

desk verdict A plausible meta-representation architecture for load forecasting with large reported gains, but the experiments don't isolate the mechanism and the 'never worse' guarantee isn't proven. read the letter →

arxiv 2506.23201 v1 pith:66OQFOFC submitted 2025-06-29 cs.LG cs.SYeess.SY

Haoran Li , Muhao Guo , Marija Ilic , Yang Weng , Guangchun Ruan This is my paper
classification cs.LGcs.SYeess.SY
keywords residentialloadforecastingmeta-representationmixture-of-expertsprobabilistichypernetworkexternaldatadeeplearningrobustness
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's central claim is that external information—weather, calendar flags, prices—should not be fed to a load-forecasting model as extra input features but should instead act as meta-knowledge that dynamically modifies the model's own parameters at each time step. To instantiate this, the authors build M2oE2, a meta mixture of experts for external data: a set of hypernetwork experts turns each external signal into a weight matrix, a gating network selects the most relevant experts, and the weighted result replaces a fixed input projection in a base GRU forecaster. On five residential and solar load datasets, they report test MSE and CRPS at roughly 5-50 percent of the values for ARIMA, CNN-GRU, RNN-GRU, LSTM, and Informer. If the claim holds, grid operators gain a plug-and-play way to make any sequence forecaster context-aware without redesigning the model. The paper also claims a residual connection guarantees the meta-module never underperforms the base forecaster.

What carries the argument

The load-bearing object is the time-varying input-projection matrix $\theta_i$, generated by a mixture of hypernetwork experts: $\theta_i = \sum_{j=1}^M l_{ji} g_j(w_{ji}) + \theta_0$, with $g_j$ mapping external source $w_{ji}$ to a candidate weight matrix and $l_{ji}$ a top-$m$ softmax gate. The matrix is applied as $x'_i = \theta_i x_i$ before the recurrent update, so the external context changes how the model reads its input rather than adding an input channel. Layer normalization on each $g_j$ output handles the heterogeneity of continuous and categorical external data, and the residual $\theta_0$ gives the claimed fallback to the base forecaster.

What would settle it

Run the same 4-layer GRU on the five datasets under identical training, in three variants: without the meta-layer, with external data concatenated to the input, and with the full M2oE2 wrapper; if either non-meta variant matches M2oE2's MSE and CRPS within noise, the meta-representation claim fails. Also check whether the gating softmax can ever output an exactly zero expert contribution—if it cannot, the claimed 'never worse' residual guarantee is not a mathematical guarantee.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is a shift in how external covariates are used: rather than concatenating temperature or calendar labels with load history, M2oE2 treats them as meta-knowledge that rewires the forecaster. Mathematically, at time step i each external source $w_{ji}$ is passed through a hypernetwork $g_j$, layer-normalized, and combined through a top-$m$ softmax gate into a time-varying matrix $\theta_i = \sum_j l_{ji} g_j(w_{ji}) + \theta_0$. The base GRU then receives $x'_i = \theta_i x_i$ instead of $x_i$, so external conditions control how the input is projected into the hidden state. With $\theta_0$ as a static residual term, the design claims a worst-case guarantee: if the external experts contribute nothing useful, the model falls back to a fixed projection and should not underperform the base network. The paper also extends the base model to probabilistic forecasting with a latent variable and ELBO training, producing predictive mean and variance, and reports that M2oE2 dominates all baselines on MSE and CRPS across all five datasets.

Load-bearing premise

The load-bearing premise is that the reported accuracy gains come from treating external data as meta-knowledge, but the experiments never compare M2oE2 against the same base GRU without the meta-layer or with external data simply concatenated, and no variance or significance tests are reported.

Editorial extensions

If this is right

  • If the central claim is right, any sequence forecasting model can be made context-aware by wrapping its input in an M2oE2-style projection, with no change to the recurrent core.
  • External signals that are slowly varying, such as temperature or day type, gain influence because even a small change in $w_{ji}$ can produce a large change in $\theta_i$.
  • The top-$m$ gating gives a built-in robustness mechanism: redundant or uninformative external sources are down-weighted at each time step, which should reduce overfitting as the number of external data streams grows.
  • The claimed residual guarantee implies that adding the meta-module should never hurt accuracy relative to the base network, so operators can adopt it without a safety trade-off.
  • In the probabilistic variant, Monte Carlo sampling from a latent variable yields predictive distributions, so the same framework covers both point and interval forecasts.

Reading between the lines

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

  • Because M2oE2 only replaces the input projection, the same wrapper could be attached to non-recurrent forecasters such as Transformers or state-space models; the paper only demonstrates it with a GRU base.
  • The gate is written as $l(h_i)$ in Eq. (12), but $h_i$ is produced from the $\theta_i$ that the gate is building, making the computation circular as written; a consistent reading uses the previous hidden state $h_{i-1}$, and the paper does not flag this ambiguity.
  • A simple ablation—same GRU with no meta-layer, and same GRU with external data concatenated—would quantify how much of the reported 5-50 percent error ratio comes from the meta-representation mechanism itself versus added parameters; the paper does not include this comparison.
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

4 major / 5 minor

Summary. The paper proposes M2oE2, a probabilistic load forecasting framework that treats external data (temperature, day type, season) as meta-knowledge rather than as direct inputs. A hypernetwork generates candidate parameter modulations from each external source, a sparse Mixture-of-Experts gate combines them, and a residual connection adds a static trainable matrix. The modulated matrix is applied to the input load vector before a 4-layer GRU encoder-decoder produces a Gaussian predictive distribution, trained with an ELBO objective. Experiments on five load/solar datasets compare MSE and CRPS with ARIMA, CNN-GRU, RNN-GRU, LSTM, and Informer. The central claims are that M2oE2 consistently outperforms all baselines in point accuracy, sharpness, and calibration, that its error is 5% to 50% of the baselines, and that the residual architecture gives a worst-case guarantee that M2oE2 never performs worse than the base DL model.

Significance. The core idea—using external factors to modulate model parameters rather than concatenating them as features—is a plausible and practically relevant architectural direction for load forecasting. The framework is modular, plug-and-play with respect to the base model, and the authors provide public code and datasets, which is a clear strength. If the empirical claims were supported by controlled ablations and uncertainty quantification, the paper would be a useful contribution to the load-forecasting literature. As it stands, however, the evidence does not yet isolate the proposed mechanism as the source of the reported gains, and the asserted worst-case guarantee is not established.

major comments (4)
  1. [Section V-C, Tables II and III] The central empirical claim of consistent superiority rests on single reported test MSE and CRPS values per method and dataset, with no standard deviations, no multiple seeds, and no significance tests. Given the large reported margins (5% to 50% of baseline error), the authors should report means and variances over several random initializations and, if possible, paired significance tests. Without this, the claim of "consistently outperforming" is not quantitatively supported.
  2. [Section V-B and V-C] There is no ablation against the base model itself: the 4-layer GRU used inside M2oE2 is never trained alone, nor with the same external features concatenated directly to its input. This is a load-bearing omission because the reported gains could plausibly come from the larger input dimension (dx' = 40 versus dx = 1 in Eq. (15)), from better hyperparameter choices, or from the base GRU's intrinsic strength, rather than from the meta-representation mechanism. Adding these two controls is necessary to attribute the improvements to the proposed idea.
  3. [Section IV-B, Eq. (14)] The claimed worst-case guarantee that M2oE2 "will never perform worse than the DL model" does not follow from Eq. (14). When the external contribution vanishes, theta_i degenerates to theta_0, but theta_0 is a static trainable matrix, not necessarily equal to the original input-to-hidden weight of the base model; moreover, the base model continues to be trained jointly, so a skip connection does not formally ensure that the learned M2oE2 solution dominates the learned base-model solution. The text should be revised to state this as an inductive bias or should provide explicit conditions under which the guarantee actually holds.
  4. [Section III-D and IV-B, Eqs. (12) and (14), Fig. 2] The gating input is specified inconsistently. The text and equations state that the gate consumes h_i, but h_i is computed only after the modulated parameter theta_i is applied in Eq. (1), and theta_i itself depends on l(h_i), creating a circular dependency. Fig. 2 indicates h_{i-1} as the gating input, which would resolve the ambiguity. The authors should correct the notation and explicitly state the temporal ordering of the computation.
minor comments (5)
  1. [Section V-B] The sentence "we have M = 3 experts (theta_i)" is imprecise: each expert is the hypernetwork output g_j(w_ji), while theta_i is the weighted sum in Eq. (14). Please rephrase.
  2. [Section IV-A, Table I] The notation "N·M·dx·dh" in Table I is unclear; N and M appear to be counts over time and external sources, but the table is supposed to summarize parameter sizes. Clarify whether these are per-instance sizes or total sizes over the dataset.
  3. [Section V-C] The phrase "average test MSE and CRPS" suggests multiple runs, but only single numbers are reported. If the tables are from one run, state so explicitly and add the requested variance information.
  4. [Equations (12)-(14)] In Eq. (13), the indicator I(j in T_i) makes gating values for non-selected experts zero, but the softmax denominator is only over T_i; this is correct but should be stated more explicitly to avoid confusion with standard softmax.
  5. [Abstract and Section VI] The claim that error is "5% to 50% of the error" of all other methods should be qualified: it holds on the particular datasets and configurations tested, and the comparison set does not include a direct base-model ablation.

Circularity Check

1 steps flagged · score 1.0 of 10

No significant circularity; the only circular step is a self-referential gating definition (l(h_i) vs h_{i-1}). The empirical claims are not circular but lack ablations.

  1. self definitional [Section IV-B, Eqs. (12), (14), (15); Section III-D]
    "The produced meta experts are filtered by a gating network l(hi), shown in the green box in Fig. 2. ... The output hi of the base DL model in turn works as input to the gating neural network in Eq. (14)."

    The recurrent state h_i is computed from the current input; in M2oE2 the input is x'_i = theta_i x_i (Eq. (15)), where theta_i = sum_j l_ji g_j(w_ji) + theta_0 and l_ji = l(h_i)[j] (Eqs. (12) and (14)). Thus theta_i depends on h_i at the same time step, while h_i depends on theta_i. This makes the forward pass self-referential and not evaluable as written. Fig. 2 draws the gate input from h_{i-1}, but the text never flags the discrepancy, so the definition is circular unless read with h_{i-1}.

full rationale

The central empirical claim that M2oE2 outperforms baselines on held-out test sets is obtained by standard ELBO/MSE training on train/test splits; there is no fitting of constants to the test set and no reduction of a reported metric to an input by construction. The residual 'never worse' guarantee is asserted rather than proven because theta_0 is a trainable weight matrix and the base model is also retrained, but this is a proof gap, not circularity. The only genuine circular step is the gating-input definition: the text specifies l(h_i), yet h_i itself is produced using theta_i, which depends on l(h_i). This is likely a typo given Fig. 2 shows h_{i-1}, but as written it is a self-referential definition. The self-citation of prior work [32] is background and not load-bearing. The absence of ablations against the base GRU alone and the lack of variance/significance reporting are evidentiary limitations affecting the strength of the empirical claim, but they do not constitute circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard deep learning components (GRU, hypernetwork, MoE, VAE) and on dataset assumptions. The only hand-tuned hyperparameters are listed above; no new physical entities or fitted fundamental constants are introduced. The key unexamined assumption is the gating input dependency, which affects the entire forward pass.

free parameters (6)
  • KL weight lambda = 0.01
    Weights the KL term in ELBO (Eq. (11)); chosen by hand, no sensitivity analysis reported.
  • top-m active experts = 2 (of M=3)
    Sparsity level in MoE gate (Eq. (13)); chosen without sensitivity analysis.
  • base GRU hidden units dh = 64
    Capacity of base sequence model; chosen by hand.
  • latent dimension dz = 32
    Variational latent size in decoder; chosen by hand.
  • meta-network hidden units = 40 = dx*dx'
    Dimension of input transform; chosen by hand.
  • training epochs and learning rate = 300, 1e-3
    Optimization choices; no early stopping described.
assumptions (4)
  • standard math Variational inference with reparameterization is a valid training objective (Eqs. (6)-(11)).
    Standard VAE machinery, not proved in the paper.
  • domain assumption All external sources w_ji are available and time-aligned for every forecast step.
    Datasets are treated as complete; no missing-data handling is described.
  • domain assumption The closed-form Gaussian CRPS (Eq. (17)) adequately represents the true predictive distribution.
    The predictive distribution from MC sampling (Eq. (10)) is not exactly Gaussian, yet the CRPS formula assumes Gaussianity.
  • ad hoc to paper The model is computable without fixed-point iteration; the gating input must be h_{i-1}, not h_i.
    The text in Eq. (12) and Section IV-B says l(h_i), but h_i depends on θ_i; only h_{i-1} is available before computing θ_i, and the paper does not clarify this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of External Data-Enhanced Meta-Representation for Adaptive Probabilistic Load Forecasting." pith.science (2026). https://pith.science/paper/66OQFOFC

@misc{pith2026250623201,
  author       = {Pith},
  title        = {Pith review of: External Data-Enhanced Meta-Representation for Adaptive Probabilistic Load Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/66OQFOFC}},
  note         = {Machine review of arXiv:2506.23201}
}
read the original abstract

Accurate residential load forecasting is critical for power system reliability with rising renewable integration and demand-side flexibility. However, most statistical and machine learning models treat external factors, such as weather, calendar effects, and pricing, as extra input, ignoring their heterogeneity, and thus limiting the extraction of useful external information. We propose a paradigm shift: external data should serve as meta-knowledge to dynamically adapt the forecasting model itself. Based on this idea, we design a meta-representation framework using hypernetworks that modulate selected parameters of a base Deep Learning (DL) model in response to external conditions. This provides both expressivity and adaptability. We further integrate a Mixture-of-Experts (MoE) mechanism to enhance efficiency through selective expert activation, while improving robustness by filtering redundant external inputs. The resulting model, dubbed as a Meta Mixture of Experts for External data (M2oE2), achieves substantial improvements in accuracy and robustness with limited additional overhead, outperforming existing state-of-the-art methods in diverse load datasets. The dataset and source code are publicly available at https://github.com/haorandd/M2oE2\_load\_forecast.git.

Figures

Figures reproduced from arXiv: 2506.23201 by the authors.

Figure 1
Figure 1. The visualization of normalized external data and loads. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The proposed M2oE2 framework. DL model encodes input load data into hidden states. To select an effective candidate while maintaining manageable computational demands, we make the following observations: • Generality: Since we aim to support both deterministic and probabilistic forecasting, parameters associated with the latent variable zi+1 are excluded from consideration. • Efficiency: In practice, dx < dh. The in… view at source ↗
Figure 3
Figure 3. The forecast and the true loads in week 1. ARIMA CNNGRU LSTM Informer RNN 𝐌 𝟐 𝐨 𝐄 𝟐 Hour Hour Hour Hour Hour Ashrae Building Load Spain Load Tetouan Load Houston Residential Load Kaggle Solar Value Value Value Value Value Value [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The forecast and the true loads in week 3 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 43 canonical work pages

  1. [1]

    A task-based day-ahead load forecasting model for stochastic economic dispatch,

    J. Han, L. Yan, and Z. Li, “A task-based day-ahead load forecasting model for stochastic economic dispatch,” IEEE Transactions on Power Systems, vol. 36, no. 6, pp. 5294–5304, 2021

    work page 2021

  2. [2]

    Deep learning-based short-term load forecasting for sup- porting demand response program in hybrid energy system,

    S. H. Pramono, M. Rohmatillah, E. Maulana, R. N. Hasanah, and F. Hario, “Deep learning-based short-term load forecasting for sup- porting demand response program in hybrid energy system,” Energies, vol. 12, no. 17, p. 3359, 2019

    work page 2019

  3. [3]

    Intelligent load forecasting and renew- able energy integration for enhanced grid reliability,

    A. Saxena, R. Shankar, E. El-Saadany, M. Kumar, O. Al Zaabi, K. Al Hosani, and U. R. Muduli, “Intelligent load forecasting and renew- able energy integration for enhanced grid reliability,” IEEE Transactions on Industry Applications , 2024

    work page 2024

  4. [4]

    Domain adaptation in physical systems via graph kernel,

    H. Li, H. Tong, and Y . Weng, “Domain adaptation in physical systems via graph kernel,” in Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , ser. KDD ’22. New York, NY , USA: Association for Computing Machinery, 2022, p. 868–876. [Online]. Available: https://doi.org/10.1145/3534678.3539380

    arXiv 2022

  5. [5]

    A transfer learning framework for power system event identification,

    H. Li, Z. Ma, and Y . Weng, “A transfer learning framework for power system event identification,” IEEE Transactions on Power Systems , vol. 37, no. 6, pp. 4424–4435, 2022

    work page 2022

  6. [6]

    Structural tensor learning for event identification with limited labels,

    H. Li, Z. Ma, Y . Weng, E. Blasch, and S. Santoso, “Structural tensor learning for event identification with limited labels,” IEEE Transactions on Power Systems, vol. 38, no. 6, pp. 5314–5328, 2023

    work page 2023

  7. [7]

    A novel sequence to sequence data modelling based cnn-lstm algorithm for three years ahead monthly peak load forecasting,

    X. Zhang, T. K. Chau, Y . H. Chow, T. Fernando, and H. H.-C. Iu, “A novel sequence to sequence data modelling based cnn-lstm algorithm for three years ahead monthly peak load forecasting,” IEEE Transactions on Power Systems, vol. 39, no. 1, pp. 1932–1947, 2023

    work page 1932

  8. [8]

    A data-driven bottom-up approach for spatial and temporal electric load forecasting,

    C. Ye, Y . Ding, P. Wang, and Z. Lin, “A data-driven bottom-up approach for spatial and temporal electric load forecasting,” IEEE Transactions on Power Systems, vol. 34, no. 3, pp. 1966–1979, 2019

    work page 1966

Show all 50 references
  1. [9]

    A novel cnn-gru-based hybrid approach for short-term residential load forecasting,

    M. Sajjad, Z. A. Khan, A. Ullah, T. Hussain, W. Ullah, M. Y . Lee, and S. W. Baik, “A novel cnn-gru-based hybrid approach for short-term residential load forecasting,” Ieee Access , vol. 8, pp. 143 759–143 768, 2020

    2020

  2. [10]

    Forecasting demand flexibility of aggregated residential load using smart meter data,

    J. Pono ´cko and J. V . Milanovi ´c, “Forecasting demand flexibility of aggregated residential load using smart meter data,” IEEE Transactions on Power Systems, vol. 33, no. 5, pp. 5446–5455, 2018

    2018

  3. [11]

    Short-term residential load forecasting based on lstm recurrent neural network,

    W. Kong, Z. Y . Dong, Y . Jia, D. J. Hill, Y . Xu, and Y . Zhang, “Short-term residential load forecasting based on lstm recurrent neural network,” IEEE transactions on smart grid , vol. 10, no. 1, pp. 841–851, 2017

    2017

  4. [12]

    Review of smart meter data analytics: Applications, methodologies, and challenges,

    Y . Wang, Q. Chen, T. Hong, and C. Kang, “Review of smart meter data analytics: Applications, methodologies, and challenges,” IEEE Transactions on smart Grid , vol. 10, no. 3, pp. 3125–3148, 2018

    2018

  5. [13]

    Short-term load forecasting using general exponential smoothing,

    W. Christiaanse, “Short-term load forecasting using general exponential smoothing,” IEEE Transactions on Power Apparatus and Systems, no. 2, pp. 900–911, 1971

    1971

  6. [14]

    Probabilistic baseline estimation via gaus- sian process,

    Y . Weng and R. Rajagopal, “Probabilistic baseline estimation via gaus- sian process,” in 2015 IEEE Power & Energy Society General Meeting . IEEE, 2015, pp. 1–5

    2015

  7. [15]

    An adaptive communication scheme for bandwidth limited residential load forecasting,

    G. Xie, X. Chen, and Y . Weng, “An adaptive communication scheme for bandwidth limited residential load forecasting,” in 2017 North American Power Symposium (NAPS). IEEE, 2017, pp. 1–6

    2017

  8. [16]

    An integrated gaussian process modeling framework for resi- dential load prediction,

    ——, “An integrated gaussian process modeling framework for resi- dential load prediction,” IEEE Transactions on Power Systems , vol. 33, no. 6, pp. 7238–7248, 2018

    2018

  9. [17]

    Probabilistic baseline estimation based on load patterns for better residential customer rewards,

    Y . Weng, J. Yu, and R. Rajagopal, “Probabilistic baseline estimation based on load patterns for better residential customer rewards,” Inter- national Journal of Electrical Power & Energy Systems , vol. 100, pp. 508–516, 2018

    2018

  10. [18]

    Input modeling and uncertainty quan- tification for improving volatile residential load forecasting,

    G. Xie, X. Chen, and Y . Weng, “Input modeling and uncertainty quan- tification for improving volatile residential load forecasting,” Energy, vol. 211, p. 119007, 2020

    2020

  11. [19]

    Short term load forecasting using xgboost,

    R. A. Abbasi, N. Javaid, M. N. J. Ghuman, Z. A. Khan, S. Ur Rehman, and Amanullah, “Short term load forecasting using xgboost,” in Web, artificial intelligence and network applications: proceedings of the workshops of the 33rd international conference on advanced information n...

    2019

  12. [20]

    Short-term load forecasting for power systems with high-penetration renewables based on multivariate data slicing transformer neural network,

    W. Lu and X. Chen, “Short-term load forecasting for power systems with high-penetration renewables based on multivariate data slicing transformer neural network,” Frontiers in Energy Research , vol. 12, p. 1355222, 2024

    2024

  13. [21]

    Recurrent neural networks for multivariate time series with missing values,

    Z. Che, S. Purushotham, K. Cho, D. Sontag, and Y . Liu, “Recurrent neural networks for multivariate time series with missing values,” Scientific reports, vol. 8, no. 1, p. 6085, 2018

    2018

  14. [22]

    Enhance load forecastability: Optimize data sampling policy by reinforcing user behaviors,

    G. Xie, X. Chen, and Y . Weng, “Enhance load forecastability: Optimize data sampling policy by reinforcing user behaviors,” European Journal of Operational Research , vol. 295, no. 3, pp. 924–934, 2021

    2021

  15. [23]

    Transformer-based model for electrical load forecasting,

    A. L’Heureux, K. Grolinger, and M. A. Capretz, “Transformer-based model for electrical load forecasting,” Energies, vol. 15, no. 14, p. 4993, 2022

    2022

  16. [24]

    Ultra-short-term load forecasting based on convolutional-lstm hybrid networks,

    H. Dong, J. Zhu, S. Li, T. Luo, H. Li, and Y . Huang, “Ultra-short-term load forecasting based on convolutional-lstm hybrid networks,” in 2022 IEEE 31st International Symposium on Industrial Electronics (ISIE) . IEEE, 2022, pp. 142–148

    2022

  17. [25]

    Short-term residential load forecasting based on the fusion of customer load uncertainty feature extraction and meteorological factors,

    W. Cao, H. Liu, X. Zhang, Y . Zeng, and X. Ling, “Short-term residential load forecasting based on the fusion of customer load uncertainty feature extraction and meteorological factors,” Sustainability, vol. 17, no. 3, p. 1033, 2025

    2025

  18. [26]

    Uncovering dominant features in short-term power load forecasting based on multi-source feature,

    P. Zeng, M. F. Elahe, J. Xu, and M. Jin, “Uncovering dominant features in short-term power load forecasting based on multi-source feature,” arXiv preprint arXiv:2103.12534 , 2021

    2021 arXiv

  19. [27]

    Physics- informed neural networks for building thermal modeling and demand response control,

    Y . Chen, Q. Yang, Z. Chen, C. Yan, S. Zeng, and M. Dai, “Physics- informed neural networks for building thermal modeling and demand response control,” Building and Environment, vol. 234, p. 110149, 2023

    2023

  20. [28]

    A physics-informed deep learning paradigm for transient power angle stability assessment,

    X. Li, S. Chen, J. Zhang, J. Gao, and Y . Bai, “A physics-informed deep learning paradigm for transient power angle stability assessment,” IEEE Journal of Radio Frequency Identification , vol. 6, pp. 948–952, 2022

    2022

  21. [29]

    A physics-informed dynamic deep au- toencoder for accurate state-of-health prediction of lithium-ion battery,

    Z. Xu, Y . Guo, and J. H. Saleh, “A physics-informed dynamic deep au- toencoder for accurate state-of-health prediction of lithium-ion battery,” Neural Computing and Applications, vol. 34, no. 18, pp. 15 997–16 017, 2022

    2022

  22. [30]

    Hypernetworks,

    D. Ha, A. Dai, and Q. V . Le, “Hypernetworks,” arXiv preprint arXiv:1609.09106, 2016

    2016 arXiv

  23. [31]

    On layer normalization in the transformer architecture,

    R. Xiong, Y . Yang, D. He, K. Zheng, S. Zheng, C. Xing, H. Zhang, Y . Lan, L. Wang, and T. Liu, “On layer normalization in the transformer architecture,” in International conference on machine learning. PMLR, 2020, pp. 10 524–10 533

    2020

  24. [32]

    Exarnn: An environment-driven adaptive rnn for learning non-stationary power dy- namics,

    H. Li, M. Guo, Y . Weng, M. Ilic, and G. Ruan, “Exarnn: An environment-driven adaptive rnn for learning non-stationary power dy- namics,” arXiv preprint arXiv:2505.17488 , 2025

    2025 arXiv

  25. [33]

    Mixture-of-experts with expert choice routing,

    Y . Zhou, T. Lei, H. Liu, N. Du, Y . Huang, V . Zhao, A. M. Dai, Q. V . Le, J. Laudon et al. , “Mixture-of-experts with expert choice routing,” Advances in Neural Information Processing Systems , vol. 35, pp. 7103– 7114, 2022

    2022

  26. [34]

    Few-shot load forecasting under data scarcity in smart grids: A meta-learning approach,

    G. Tsoumplekas, C. Athanasiadis, D. I. Doukas, A. Chrysopoulos, and P. Mitkas, “Few-shot load forecasting under data scarcity in smart grids: A meta-learning approach,” Energies, vol. 18, no. 3, p. 742, 2025

    2025

  27. [35]

    Metaeformer: Unveiling and leveraging meta-patterns for complex and dynamic systems load forecasting,

    S. Huang, T. Zhang, Z. Zhang, X. Wang, L. Wang, and X. Wang, “Metaeformer: Unveiling and leveraging meta-patterns for complex and dynamic systems load forecasting,” arXiv preprint arXiv:2506.12800 , 2025

    2025 arXiv

  28. [36]

    What uncertainties do we need in bayesian deep learning for computer vision?

    A. Kendall and Y . Gal, “What uncertainties do we need in bayesian deep learning for computer vision?” Advances in neural information processing systems, vol. 30, 2017

    2017

  29. [37]

    Stochastic backprop- agation and approximate inference in deep generative models,

    D. J. Rezende, S. Mohamed, and D. Wierstra, “Stochastic backprop- agation and approximate inference in deep generative models,” in International conference on machine learning. PMLR, 2014, pp. 1278– 1286

    2014

  30. [38]

    Learning feature engineering for classification

    F. Nargesian, H. Samulowitz, U. Khurana, E. B. Khalil, and D. S. Turaga, “Learning feature engineering for classification.” in Ijcai, vol. 17, 2017, pp. 2529–2535

    2017

  31. [39]

    Neural networks for machine learning lecture 6a overview of mini-batch gradient descent,

    G. Hinton, N. Srivastava, and K. Swersky, “Neural networks for machine learning lecture 6a overview of mini-batch gradient descent,” Cited on, vol. 14, no. 8, p. 2, 2012

    2012

  32. [40]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

    2016

  33. [41]

    Great Energy Predictor III,

    Addison Howard, Chris Balbach, Clayton Miller, Jeff Haberl, Krishnan Gowri, and Sohier Dane, “Great Energy Predictor III,” 2019. [Online]. Available: https://www.kaggle.com/competitions/ ashrae-energy-prediction/data

    2019

  34. [42]

    Hourly energy demand generation and weather,

    Kolasniwash, “Hourly energy demand generation and weather,” 2019. [Online]. Available: https://www.kaggle.com/datasets/nicholasjhana/ energy-consumption-generation-prices-and-weather

    2019

  35. [43]

    Electric Power Consumption,

    Fedesoriano, “Electric Power Consumption,” 2022. [On- line]. Available: https://www.kaggle.com/datasets/fedesoriano/ electric-power-consumption

    2022

  36. [44]

    Residential power usage 3years data - Timeseries,

    Sri Polu, “Residential power usage 3years data - Timeseries,”

  37. [45]

    Renewable Energy and Weather Conditions,

    AI Maverick, “Renewable Energy and Weather Conditions,” 2023. [Online]. Available: https://www.kaggle.com/datasets/samanemami/ renewable-energy-and-weather-conditions

    2023

  38. [46]

    Short-term stochastic load forecasting using autoregressive integrated moving average models and hidden markov model,

    J. P. Hermias, K. Teknomo, and J. C. N. Monje, “Short-term stochastic load forecasting using autoregressive integrated moving average models and hidden markov model,” in 2017 international conference on infor- mation and communication technologies (ICICT) . IEEE, 2017, pp. 131–137

    2017

  39. [47]

    Short-term power load forecasting of residential community based on gru neural network,

    J. Zheng, X. Chen, K. Yu, L. Gan, Y . Wang, and K. Wang, “Short-term power load forecasting of residential community based on gru neural network,” in 2018 International conference on power system technology (POWERCON). IEEE, 2018, pp. 4862–4868

    2018

  40. [48]

    Informer: Beyond efficient transformer for long sequence time-series forecasting,

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, “Informer: Beyond efficient transformer for long sequence time-series forecasting,” in Proceedings of the AAAI conference on artificial intel- ligence, vol. 35, no. 12, 2021, pp. 11 106–11 115

    2021

  41. [49]

    Decomposition of the continuous ranked probability score for ensemble prediction systems,

    H. Hersbach, “Decomposition of the continuous ranked probability score for ensemble prediction systems,” Weather and Forecasting , vol. 15, no. 5, pp. 559–570, 2000

    2000

  42. [2019]

    Available: https://www.kaggle.com/datasets/srinuti/ residential-power-usage-3years-data-timeseries 10

    [Online]. Available: https://www.kaggle.com/datasets/srinuti/ residential-power-usage-3years-data-timeseries 10

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.