Pith. sign in

REVIEW 5 major objections 7 minor 2 cited by

ExARNN: An Environment-Driven Adaptive RNN for Learning Non-Stationary Power Dynamics

T0 review · 5 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read ExARNN feeds weather through an NCDE to generate an RNN's recurrence weights, reaching 1.82% MAPE on Spain and 4.64% on Texas, beating four baselines.

desk verdict A clean idea—hypernetwork + NCDE for weather-adaptive RNN weights—but the empirical case is underbuilt and one unstated spline detail could invalidate the headline numbers. read the letter →

arxiv 2505.17488 v1 pith:6UIKLGF7 submitted 2025-05-23 cs.LG cs.SYeess.SY

Haoran Li , Muhao Guo , Yang Weng , Marija Ilic , Guangchun Ruan This is my paper
classification cs.LGcs.SYeess.SY
keywords non-stationaryloadforecastinghypernetworkneuralcontrolleddifferentialequationsadaptiveRNNweather-drivenparametergenerationirregulartimeseriespowersystemdynamics
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

ExARNN is a recurrent model whose recurrence weights are generated, at every time step, by a neural controlled differential equation driven by external measurements such as temperature. The paper's central claim is that this environment-driven parameter adaptation lets a base RNN track non-stationary power load dynamics better than fixed-weight RNNs or ODE-based baselines. On the Spain and Texas datasets, the authors report MAPE of 1.82% and 4.64%, respectively, the lowest among RNN, RNN-$\Delta t$, ODE-RNN, and NCDE. A sympathetic reader would care because load and renewable generation are increasingly weather-driven, and the method offers a way to fuse low-frequency weather data with high-frequency power measurements without manual alignment, with the stated goal of sample efficiency and generalization to unseen weather patterns.

What carries the argument

The load-bearing object is the environment-driven parameter generator: an NCDE $\hat{h}$ that, for each power timestamp $t_i$, evaluates the feature flow $z(t_i, W) = z(t_1) + \int_{t_1}^{t_i} \hat{h}(z(s)) \frac{dW}{ds}(s)\,ds$, where $W(t)$ is a natural cubic spline through the weather measurements augmented with time, and then applies a mapping $l_2$ to produce the recurrent weight $\theta_1(t_i, W)$ of the main RNN. This carries the argument because it turns sparse, irregularly sampled weather into a continuously available control signal that can be evaluated at arbitrary power timestamps, and it makes the base RNN's dynamics environment-adaptive rather than merely environment-augmented. The static parameters $\theta_0$ and the generator parameters $\psi$ are the only trainable quantities, which keeps training end-to-end and sample-efficient.

What would settle it

Recompute the natural cubic spline $W(t)$ at each test step using only weather measurements up to that step (or use a one-sided smoother), retrain ExARNN, and compare its MAPE against RNN-$\Delta t$ and ODE-RNN on the same splits; if the gap shrinks or reverses, the reported superiority depends on future weather leakage. A complementary probe: perturb a weather measurement after the prediction time and check whether the forecast at an earlier time changes, which would confirm leakage through the global spline.

Watch

Extended reading notes

Core claim

The paper claims that external data can be treated as meta-knowledge: instead of concatenating weather features to the input, a hypernetwork maps a continuous weather feature flow to the recurrent weight matrix $\theta_1(t, W)$ of a main RNN, so the recurrence dynamically decides how much past state to carry forward depending on the environment. The continuous flow is produced by an NCDE over a natural cubic spline path $W(t)$ built from sparse weather observations, which lets the model evaluate environment features at every power-system timestamp even when weather and power are sampled at different rates. Trained end-to-end with MSE loss by updating only the NCDE parameters and the static part of the RNN, ExARNN is reported to outperform all baselines on both datasets, with the largest margin on Spain (1.82% MAPE versus 4.95% for the best baseline, RNN-$\Delta t$).

Load-bearing premise

The evaluation assumes the continuous weather path used at each prediction time contains only weather observed before that prediction; the paper does not state that the cubic spline is causally masked, and a global spline would leak future weather into test forecasts.

Editorial extensions

If this is right

  • On the two evaluated datasets, ExARNN's reported MAPE of 1.82% (Spain) and 4.64% (Texas) is the lowest among RNN, RNN-$\Delta t$, ODE-RNN, and NCDE, with MSE of 0.0001 and 0.0458 respectively.
  • Because weather enters through generated weights rather than as an input feature, the model should be less prone to overfitting specific weather patterns and better able to handle new conditions.
  • The NCDE-based continuous path removes the need to manually align 60-minute weather data with 15-minute power data, since environment features can be evaluated at any timestamp.
  • Test-time inference of 0.038 seconds (Texas) and 0.090 seconds (Spain) per sample is fast enough for operational use despite being slower than a vanilla RNN.
  • The training scheme updates only the generator and static parameters, so the main RNN's adapted weights stay consistent with the environment at every step.

Reading between the lines

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

  • Editorial inference: the reported accuracy could partly come from a causal leak, because a natural cubic spline is a global interpolant; a one-sided or re-fit spline experiment would separate genuine adaptation from look-ahead.
  • Editorial inference: the same NCDE-generated weight mechanism could in principle adapt other model families such as LSTM or attention-based heads, but the paper only demonstrates the RNN instantiation, so that extension is untested.
  • Editorial inference: because the generator is trained end-to-end on MSE alone, nothing prevents $\theta_1(t, W)$ from varying erratically between weather observations; evaluating on held-out years or adding smoothness regularization would test whether the adaptation transfers.
  • Editorial inference: a natural next test is applying ExARNN to renewable generation or price forecasting, where weather drives dynamics even more directly; the paper does not report such experiments.
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

5 major / 7 minor

Summary. The paper proposes ExARNN, an RNN whose recurrent transition weights are generated by a hypernetwork driven by external weather and time data. The hypernetwork is an NCDE that creates a continuous feature flow by integrating against a natural cubic spline of the external measurements, which lets the model evaluate RNN parameters at the arbitrary timestamps of the higher-frequency power load data. The authors report experiments on Spain and Texas load data, claiming the lowest MAPE and MSE against RNN, RNN-dt, ODE-RNN, and NCDE baselines (Table I), with modest test-time overhead (Table II). The central contribution is the hierarchical hypernetwork-NCDE design for fusing irregularly sampled external covariates into a recurrent predictor.

Significance. The modeling idea is timely and sensible: treating weather as meta-knowledge that modulates the RNN's recurrent parameters, rather than as an extra input channel, is a plausible strategy for non-stationary load forecasting, and the NCDE formulation is a principled way to bridge 15-minute load data with 60-minute weather data. The architecture is internally consistent (Eqs. (1)-(4)), the training algorithm is simple and coherent, and the two-dataset evaluation targets a practically important problem. However, the paper's central empirical claim rests on a single table with no error bars, no hyperparameter disclosure, no code, no normalization details, and a potentially non-causal spline construction. As it stands, the evidence is not sufficient to establish the claimed superiority; the evaluation must be made reproducible and causal before the result can be accepted.

major comments (5)
  1. [IV-A Step 1, Eq. (3)] The paper does not specify how the continuous path W(t) is constructed relative to the train/test split. A natural cubic spline is a global interpolant: its value and derivative at any s < t_i depend on all knots, including weather observations at times greater than t_i. If W(t) is fit once on the full weather record, then dW/ds in the integral of Eq. (3) encodes future weather, so the NCDE flow z(t_i, W) and the resulting RNN parameters may contain test-period information. This would invalidate the reported test MAPE/MSE in Table I. Please state explicitly whether W(t) is recomputed or masked so that only data available before each prediction is used, and report the results under that causal protocol.
  2. [V-A, Table I] No experimental uncertainty is reported: the tables show a single run with no seeds, no standard deviations, and no statistical comparison. Given the small differences between ExARNN (4.64% MAPE) and RNN-dt (4.82%) or ODE-RNN (4.75%) on Texas, the claimed superiority is not established without variance estimates or a paired significance test across multiple runs.
  3. [V-A, Table I] The NCDE baseline performs at 17.78% MAPE on Texas and 19.87% on Spain, an order of magnitude worse than the other baselines and than the NCDE component inside ExARNN. This extreme gap strongly suggests unequal tuning, a different prediction setup, or a missing aggregation step for the baseline. The manuscript must report the NCDE baseline's hyperparameters, input representation, training procedure, and prediction protocol; otherwise the comparison is not interpretable and the claimed superiority over NCDE is not credible.
  4. [V-A] The evaluation protocol is under-specified: there is no description of normalization/scaling, input window length, forecasting horizon, validation split, or how the 15-minute load data is matched with the hourly weather data. The Spain MSE of 0.0001 suggests that the load target is scaled, but this is not stated. Without these details, the metric values in Table I cannot be reproduced or compared across models and datasets.
  5. [I, V] The paper claims that ExARNN is 'highly sample-efficient' and 'generalizable' (Section I and the conclusion), but no experiment varies the amount of training data or tests distribution shift. This claim is not supported by the reported results. Either provide a sample-efficiency experiment or temper the claim to what the evidence actually shows.
minor comments (7)
  1. [II] There are typographical issues in the problem statement: '{x(ti))}' and '{w(ti)}' contain mismatched parentheses, and the set relation 'Nw ⊆ Nx' is unclear because Nx and Nw denote sets of timestamps; please state the intended meaning precisely.
  2. [III-A, Eq. (1)] The notation for the hypernetwork output is inconsistent: Eq. (1) uses θ1(w), while later sections use θ1(t_i, W). Please unify the notation to make clear that the generated parameters depend on the continuous path evaluated at time t_i.
  3. [III-B, Eq. (3)] Equation (3) writes the NCDE integral with dW/ds as a control signal; since W is a natural cubic spline, this is formally fine, but the presentation should clarify that the path is piecewise polynomial and that the derivative is taken in the sense of the interpolant.
  4. [V-A, Fig. 3] Only subfigure (a) has a legend and axis labels; subfigures (b)-(e) lack legends and clear axis labels, making visual comparison difficult. Please add consistent labels to all subfigures.
  5. [V-B, Table I] The MAPE metric is never defined in the text, and Equation (4) only defines the training loss as MSE. Please define MAPE explicitly and state the aggregation (e.g., over all test timestamps).
  6. [V-C, Table II] The test-time comparison lacks hardware/software context and implementation details, so the absolute times are not reproducible. Also, reporting only inference time does not account for the overhead of building W(t) or the NCDE state, which is part of the model's operation.
  7. [References] The Spain dataset description in Section V-A says the data is from ENTSOE and the Open Weather API as part of a personal project, but reference [28] is a general machine-learning-and-climate paper and does not appear to be the dataset source. Please cite the actual data sources.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: ExARNN's predictions are computed from external inputs via standard NCDE/hypernetwork machinery and evaluated on held-out test data against external baselines; the cubic-spline causality concern is a soundness risk, not a circularity in the derivation.

full rationale

The central claim—that ExARNN achieves the lowest MAPE and MSE on the Spanish and Texas test sets (Section V-B, Table I)—is an empirical result obtained on held-out test data after end-to-end training with the MSE loss of Equation (4). The prediction chain (weather w(t_i) → cubic spline W(t) → NCDE flow z(t,W) via Eq. (3) → θ1(t,W) = l2(z(t,W)) → main RNN Eq. (1) → x̂(t_{i+1})) contains no step in which a target quantity is defined in terms of the fitted result; the model outputs are computed from inputs through standard NCDE/hypernetwork machinery cited to external work (Kidger et al. [24]; Ha et al. [23]). The reported test MAPE/MSE are genuine predictions on the test portion (Figure 3), so no fitted parameter is renamed as a prediction. Self-citations ([3], [7], [8], [11], [18], [21]) are contextual related-work references to the authors' own prior power-system and ML papers and are not load-bearing for the ExARNN architecture or the empirical superiority claim. The safety concern identified by the reviewer—that the natural cubic spline W(t) in Section IV-A Step 1 is a global interpolant and, if fit once on the full weather record, its derivative dW/ds at s < t_i encodes future weather, potentially leaking test-period context into the NCDE flow—is a real causal-evaluation risk (a soundness issue), but it is not a circularity pattern in the defined sense: the test loads are never used to construct the inputs, and the predicted value is not defined in terms of the fitted result. Accordingly, no circular step is exhibited, and the appropriate finding is 'no significant circularity' with a low score reflecting the presence of minor non-load-bearing self-citations.

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

This is an empirical ML architecture paper with no new physical entities. The central free parameters are neural network weights fitted to data, and the most consequential unstated choice is target normalization, which makes the reported MSE uninterpretable as written.

free parameters (4)
  • NCDE hypernetwork weights psi = not reported
    Weights of l1, l2, and h-hat in Equations (2)-(3). They are trained by MSE backpropagation and generate the RNN transition matrix theta1(t,W).
  • Static RNN weights theta0 = not reported
    W1, W2, b1, b2 in Equation (1). Trained jointly with psi; dimensions and initialization are not stated.
  • Load target scaling factor = not reported
    Reported MSE of 0.0001 for Spain load in the 26-32 GW range implies normalized targets; the scaling factor is unstated and affects interpretability of every metric.
  • Experiment hyperparameters = not reported
    Learning rate eta, epochs N, hidden sizes, ODE solver and tolerance, and baseline settings are omitted, so the comparison cannot be reproduced.
assumptions (4)
  • domain assumption Natural cubic spline interpolation over sparse weather data yields a valid continuous control path W(t) for the NCDE flow.
    Section IV-A Step 1 assumes hourly weather can be smoothly interpolated to 15-minute power timestamps without harming accuracy or introducing lookahead.
  • domain assumption A first-order RNN with time-varying transition matrix theta1(w) can represent the relevant non-stationary load dynamics.
    This is the modeling capacity assumption behind Equation (1); no comparison with higher-order or attention-based sequence models is given.
  • domain assumption Temperature is the only external driver needed for the two case studies.
    The datasets use temperature-like weather only; other covariates such as humidity, wind, or calendar effects are not tested.
  • standard math The ODE solver in Equations (2)-(3) approximates the continuous flow accurately enough for downstream forecasting.
    Standard numerical integration assumption; no solver tolerance or error analysis is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ExARNN: An Environment-Driven Adaptive RNN for Learning Non-Stationary Power Dynamics." pith.science (2026). https://pith.science/paper/6UIKLGF7

@misc{pith2026250517488,
  author       = {Pith},
  title        = {Pith review of: ExARNN: An Environment-Driven Adaptive RNN for Learning Non-Stationary Power Dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6UIKLGF7}},
  note         = {Machine review of arXiv:2505.17488}
}
read the original abstract

Non-stationary power system dynamics, influenced by renewable energy variability, evolving demand patterns, and climate change, are becoming increasingly complex. Accurately capturing these dynamics requires a model capable of adapting to environmental factors. Traditional models, including Recurrent Neural Networks (RNNs), lack efficient mechanisms to encode external factors, such as time or environmental data, for dynamic adaptation. To address this, we propose the External Adaptive RNN (ExARNN), a novel framework that integrates external data (e.g., weather, time) to continuously adjust the parameters of a base RNN. ExARNN achieves this through a hierarchical hypernetwork design, using Neural Controlled Differential Equations (NCDE) to process external data and generate RNN parameters adaptively. This approach enables ExARNN to handle inconsistent timestamps between power and external measurements, ensuring continuous adaptation. Extensive forecasting tests demonstrate ExARNN's superiority over established baseline models.

Figures

Figures reproduced from arXiv: 2505.17488 by the authors.

Figure 1
Figure 1. The main framework of the proposed ExARNN. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Data of load and temperature in Spain. trical consumption, generation, pricing, and weather conditions for Spain. The consumption and generation data were obtained from ENTSOE, a public platform for Transmission System Operator (TSO) data. Weather information was sourced from the Open Weather API for the five largest cities in Spain as part of a personal project and made publicly available [PITH_FULL_IMAGE:figures/… view at source ↗
Figure 3
Figure 3. Train and test performances of Spain Load Dataset for different methods. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Enabling Scalable Topology Inference in Distribution Systems via Constrained Multi-Source Inference

    cs.AI 2026-05 reject novelty 4.0 of 10

    A locality-limited constrained inference solver that combines AMI voltage, GIS coordinates, and physical capacity/voltage constraints can reportedly recover node-to-transformer connectivity at >95% accuracy on feeders...

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

    cs.LG 2025-06 conditional novelty 4.0 of 10

    M2oE2 uses hypernetworks and a mixture-of-experts gate to let external conditions reshape a GRU's input weights, reporting large gains in load forecasting error metrics.

Reference graph

Works this paper leans on

30 extracted references · 23 canonical work pages · cited by 2 Pith papers

  1. [1]

    Continuous-time stochastic state-space modeling of non- stationary power system uncertainty: A data-driven systematic realiza- tion method,

    Y . Wan, “Continuous-time stochastic state-space modeling of non- stationary power system uncertainty: A data-driven systematic realiza- tion method,” IEEE Transactions on Power Systems , pp. 1–12, 2024

    work page 2024

  2. [2]

    A review on the selected applications of forecasting models in renewable power systems,

    A. Ahmed and M. Khalid, “A review on the selected applications of forecasting models in renewable power systems,” Renewable and Sustainable Energy Reviews , vol. 100, pp. 9–21, 2019

    work page 2019

  3. [3]

    Graph mining for classifying and localizing solar panels in distribution grids,

    M. Guo, Q. Cui, and Y . Weng, “Graph mining for classifying and localizing solar panels in distribution grids,” in 2023 Panda Forum on Power and Energy (PandaFPE). IEEE, 2023, pp. 1743–1747

    work page 2023

  4. [4]

    Temperature scenario generation for probabilistic load forecasting,

    J. Xie and T. Hong, “Temperature scenario generation for probabilistic load forecasting,” IEEE Transactions on Smart Grid , vol. 9, no. 3, pp. 1680–1687, 2018

    work page 2018

  5. [5]

    Load forecasting techniques for power system: Research challenges and survey,

    N. Ahmad, Y . Ghadi, M. Adnan, and M. Ali, “Load forecasting techniques for power system: Research challenges and survey,” IEEE Access, vol. 10, pp. 71 054–71 090, 2022

    work page 2022

  6. [6]

    Detailed hourly weather measurements for power system applications,

    F. Safdarian, M. Stevens, J. Snodgrass, and T. J. Overbye, “Detailed hourly weather measurements for power system applications,” in 2024 IEEE Texas Power and Energy Conference (TPEC) . IEEE, 2024, pp. 1–6

    work page 2024

  7. [7]

    Low-dimensional ode embedding to convert low-resolution meters into “virtual

    H. Li, Z. Ma, Y . Weng, H. Zhong, and X. Zheng, “Low-dimensional ode embedding to convert low-resolution meters into “virtual” pmus,” IEEE Transactions on Power Systems , pp. 1–13, 2024. 2015-01 2015-07 2016-01 2016-07 2017-01 2017-07 2018-01 2018-07 2019-01 26 28 30 32Power Load (×10³ MW) True Load Predicted Training Load Predicted T est Load Train/T est...

    work page 2024

  8. [8]

    Pix-gan: Enhance physics-informed estimation via generative adversarial network,

    H. Li and Y . Weng, “Pix-gan: Enhance physics-informed estimation via generative adversarial network,” in 2023 IEEE International Conference on Data Mining (ICDM) , 2023, pp. 1085–1090

    work page 2023

Show all 30 references
  1. [9]

    Hyndman, Forecasting: principles and practice

    R. Hyndman, Forecasting: principles and practice . OTexts, 2018

    2018

  2. [10]

    Short- term electricity demand forecasting with mars, svr and arima models using aggregated demand data in queensland, australia,

    M. S. Al-Musaylh, R. C. Deo, J. F. Adamowski, and Y . Li, “Short- term electricity demand forecasting with mars, svr and arima models using aggregated demand data in queensland, australia,” Advanced Engineering Informatics, vol. 35, pp. 1–16, 2018

    2018

  3. [11]

    Hd-deep-em: Deep expectation maximization for dynamic hidden state recovery using heterogeneous data,

    Z. Ma, H. Li, Y . Weng, E. Blasch, and X. Zheng, “Hd-deep-em: Deep expectation maximization for dynamic hidden state recovery using heterogeneous data,” IEEE Transactions on Power Systems , vol. 39, no. 2, pp. 3575–3587, 2024

    2024

  4. [12]

    Short-term load forecasting based on a semi-parametric additive model,

    S. Fan and R. J. Hyndman, “Short-term load forecasting based on a semi-parametric additive model,” IEEE transactions on power systems , vol. 27, no. 1, pp. 134–141, 2011

    2011

  5. [13]

    Sig2vec: Dictionary design for incipient faults in distribution systems,

    Q. Cui, Y . Weng, and M. Guo, “Sig2vec: Dictionary design for incipient faults in distribution systems,” in 2023 IEEE Power & Energy Society General Meeting (PESGM) . IEEE, 2023, pp. 1–5

    2023

  6. [14]

    Msq-biobert: Ambiguity resolution to enhance biobert medical question-answering,

    M. Guo, M. Guo, E. T. Dougherty, and F. Jin, “Msq-biobert: Ambiguity resolution to enhance biobert medical question-answering,” in Proceed- ings of the ACM Web Conference 2023 , 2023, pp. 4020–4028

    2023

  7. [15]

    Bayesian iterative prediction and lexical- based interpretation for disturbed chinese sentence pair matching,

    M. Guo, M. Guo, J. Su, J. Chen, J. Yu, J. Wang, H. Du, P. Sahu, A. A. Sharma, and F. Jin, “Bayesian iterative prediction and lexical- based interpretation for disturbed chinese sentence pair matching,” in Proceedings of the ACM on Web Conference 2024 , 2024, pp. 4618– 4629

    2024

  8. [16]

    Application of support vector machine models for forecasting solar and wind energy resources: A review,

    A. Zendehboudi, M. A. Baseer, and R. Saidur, “Application of support vector machine models for forecasting solar and wind energy resources: A review,” Journal of cleaner production , vol. 199, pp. 272–285, 2018

    2018

  9. [17]

    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

  10. [18]

    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

    2023

  11. [19]

    Convolutional lstm network: A machine learning approach for precipitation nowcasting,

    X. Shi, Z. Chen, H. Wang, D.-Y . Yeung, W.-K. Wong, and W.-c. Woo, “Convolutional lstm network: A machine learning approach for precipitation nowcasting,” Advances in neural information processing systems, vol. 28, 2015

    2015

  12. [20]

    Electricity price forecasting: A review of the state-of-the-art with a look into the future,

    R. Weron, “Electricity price forecasting: A review of the state-of-the-art with a look into the future,” International journal of forecasting, vol. 30, no. 4, pp. 1030–1081, 2014

    2014

  13. [21]

    Distribution grid topology and parameter estimation using deep-shallow neural network with physical consistency,

    H. Li, Y . Weng, V . Vittal, and E. Blasch, “Distribution grid topology and parameter estimation using deep-shallow neural network with physical consistency,” IEEE Transactions on Smart Grid, vol. 15, no. 1, pp. 655– 666, 2024

    2024

  14. [22]

    Handling renewable energy variability and uncertainty in power system operation,

    R. Bessa, C. Moreira, B. Silva, and M. Matos, “Handling renewable energy variability and uncertainty in power system operation,” Advances in Energy Systems: The Large-scale Renewable Energy Integration Challenge, pp. 1–26, 2019

    2019

  15. [23]

    Hypernetworks,

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

    2016 arXiv

  16. [24]

    Neural controlled differential equations for irregular time series,

    P. Kidger, J. Morrill, J. Foster, and T. Lyons, “Neural controlled differential equations for irregular time series,” Advances in Neural Information Processing Systems , vol. 33, pp. 6696–6707, 2020

    2020

  17. [25]

    Neural ordinary differential equations,

    R. T. Chen, Y . Rubanova, J. Bettencourt, and D. K. Duvenaud, “Neural ordinary differential equations,” Advances in neural information pro- cessing systems, vol. 31, 2018

    2018

  18. [26]

    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

  19. [27]

    T. J. Lyons, M. Caruana, and T. L ´evy, Differential equations driven by rough paths. Springer, 2007

    2007

  20. [28]

    Tackling climate change with machine learning,

    D. Rolnick, P. L. Donti, L. H. Kaack, K. Kochanski, A. Lacoste, K. Sankaran, A. S. Ross, N. Milojevic-Dupont, N. Jaques, A. Waldman- Brown et al. , “Tackling climate change with machine learning,” ACM Computing Surveys (CSUR) , vol. 55, no. 2, pp. 1–96, 2022

    2022

  21. [29]

    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

  22. [30]

    Latent ordinary differential equations for irregularly-sampled time series,

    Y . Rubanova, R. T. Chen, and D. K. Duvenaud, “Latent ordinary differential equations for irregularly-sampled time series,” Advances in neural information processing systems , vol. 32, 2019

    2019

Pith tools

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