pith. sign in

arxiv: 2605.12196 · v2 · pith:22CCPLGInew · submitted 2026-05-12 · 💻 cs.LG

ECTO: Exogenous-Conditioned Temporal Operator for Ultra-Short-Term Wind Power Forecasting

Cao Yuan , Junjun Wang This is my paper

Pith reviewed 2026-05-21 08:19 UTC · model grok-4.3

classification 💻 cs.LG
keywords wind power forecastingexogenous variablesvariable selectionphysical priorsmixture of expertstemporal forecastingdeep learningmeteorological data
0
0 comments X

The pith

ECTO improves ultra-short-term wind power forecasts by selecting relevant meteorological inputs with physical priors and refining predictions through regime-specific corrections.

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

Accurate short-term wind power predictions help grid operators balance supply and demand as renewable energy grows. The paper shows that the most useful weather measurements change across sites, conditions, and horizons, so generic mixing of inputs falls short. ECTO addresses this by first using physical knowledge to pick a compact set of exogenous variables in a hierarchical and sparse way, then routing the forecast through expert modules that apply targeted gain, bias, and horizon adjustments. Experiments across three wind farms of different sizes and climates confirm lower mean squared error than strong baselines, with the advantage increasing at longer horizons. Ablations indicate both the selection and refinement steps contribute to the gains.

Core claim

ECTO decomposes exogenous modeling into Physically-Grounded Variable Selection (PGVS), which performs hierarchical group-aware sparse selection over meteorological variables using a domain-informed physical prior and sparsemax activations to produce a condition-adaptive context, and Exogenous-Conditioned Regime Refinement (ECRR), which applies a mixture-of-experts paradigm for gain-bias calibration and horizon-specific corrections. On three wind farms with capacities from 66 to 200 MW and 11 to 13 exogenous variables, ECTO records the lowest MSE, with relative improvements of 2.2 to 5.2 percent over the strongest baseline that widen to 8.6 percent at horizon 32.

What carries the argument

Physically-Grounded Variable Selection (PGVS) module that applies a domain-informed physical prior with sparsemax activations for hierarchical sparse selection of exogenous meteorological variables, paired with Exogenous-Conditioned Regime Refinement (ECRR) using mixture-of-experts for adaptive calibration.

If this is right

  • Forecast accuracy improves across farms that differ in climate, capacity, and number of available weather measurements.
  • Gains become larger as the prediction horizon extends from short to moderate lengths.
  • Each component contributes additively, with variable selection and regime refinement each adding measurable error reduction.
  • The learned selection patterns align with physical expectations and differ sensibly by site.

Where Pith is reading between the lines

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

  • The same physical-prior approach to variable selection could transfer to forecasting tasks with other renewables such as solar power where meteorological drivers also vary by location.
  • Mixture-of-experts calibration for horizon-specific corrections may help stabilize predictions in other non-stationary time series with changing external conditions.
  • Extending the physical prior to incorporate additional atmospheric relationships could further lower errors during extreme weather events not well represented in the current sites.

Load-bearing premise

The physical prior inside the variable selection step yields selection patterns that remain useful and stable across new sites, operating conditions, and prediction horizons instead of overfitting to the three tested farms.

What would settle it

Running ECTO on a fourth wind farm with previously unseen climate, capacity, or exogenous set and observing no MSE reduction relative to the strongest baseline would indicate the claimed generalization does not hold.

Figures

Figures reproduced from arXiv: 2605.12196 by Cao Yuan, Junjun Wang.

Figure 1
Figure 1. Figure 1: Overall architecture of ECTO. The target power sequence is encoded by the [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: 24-hour continuous prediction on WF1 (day starting at sample 8976). Each [PITH_FULL_IMAGE:figures/full_fig_p025_2.png] view at source ↗
Figure 2
Figure 2. Figure 2: 24-hour continuous prediction on WF1 (day starting at sample 8976). Each [PITH_FULL_IMAGE:figures/full_fig_p027_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: 16-step prediction details across four representative operating conditions on [PITH_FULL_IMAGE:figures/full_fig_p026_3.png] view at source ↗
Figure 3
Figure 3. Figure 3: 16-step prediction details across four representative operating conditions on [PITH_FULL_IMAGE:figures/full_fig_p028_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Horizon-wise RMSE on the full WF1 test set. Each point is the RMSE at a [PITH_FULL_IMAGE:figures/full_fig_p026_4.png] view at source ↗
Figure 4
Figure 4. Figure 4: Horizon-wise RMSE on the full WF1 test set. Each point is the RMSE at a [PITH_FULL_IMAGE:figures/full_fig_p029_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: 24-hour continuous prediction on WF4 (66 MW, day starting at sample 9696). [PITH_FULL_IMAGE:figures/full_fig_p027_5.png] view at source ↗
Figure 5
Figure 5. Figure 5: 24-hour continuous prediction on WF4 (66 MW, day starting at sample 9696). [PITH_FULL_IMAGE:figures/full_fig_p030_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Per-step MSE across the 16-step prediction horizon on WF1. ECTO maintains [PITH_FULL_IMAGE:figures/full_fig_p030_6.png] view at source ↗
Figure 6
Figure 6. Figure 6: Per-step MSE across the 16-step prediction horizon on WF1. ECTO maintains [PITH_FULL_IMAGE:figures/full_fig_p033_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Global average PGVS variable weights across three wind farms. Bars are color [PITH_FULL_IMAGE:figures/full_fig_p031_7.png] view at source ↗
Figure 7
Figure 7. Figure 7: Global average PGVS variable weights across three wind farms. Bars are color [PITH_FULL_IMAGE:figures/full_fig_p034_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Raw sample-level PGVS variable-weight heatmaps. Each panel shows individual [PITH_FULL_IMAGE:figures/full_fig_p032_8.png] view at source ↗
Figure 8
Figure 8. Figure 8: Raw sample-level PGVS variable-weight heatmaps. Each panel shows individual [PITH_FULL_IMAGE:figures/full_fig_p036_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: PGVS variable-weight heatmaps averaged by power bin. Each row is an exoge [PITH_FULL_IMAGE:figures/full_fig_p033_9.png] view at source ↗
Figure 9
Figure 9. Figure 9: PGVS variable-weight heatmaps averaged by power bin. Each row is an exoge [PITH_FULL_IMAGE:figures/full_fig_p036_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: ECRR calibration strategy by dominant regime. Each point is a test sample; [PITH_FULL_IMAGE:figures/full_fig_p034_10.png] view at source ↗
Figure 10
Figure 10. Figure 10: ECRR calibration strategy by dominant regime. Each point is a test sample; [PITH_FULL_IMAGE:figures/full_fig_p038_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Autocorrelation functions of the Diebold-Mariano loss differential series (ECTO [PITH_FULL_IMAGE:figures/full_fig_p046_11.png] view at source ↗
read the original abstract

Accurate ultra-short-term wind power forecasting is critical for grid dispatch and reserve management, yet remains challenging due to the non-stationary, condition-dependent nature of wind generation. Meteorological exogenous variables carry substantial predictive information, but the most informative variable combination varies across sites, operating conditions, and prediction horizons. Existing deep learning approaches either treat exogenous inputs as generic auxiliary channels through uniform mixing or soft gating, or rely on fixed preprocessing steps such as PCA, without exploiting the physical structure of meteorological variables. We propose ECTO (Exogenous-Conditioned Temporal Operator), a unified framework that decomposes exogenous variable modeling into two complementary modules. Physically-Grounded Variable Selection (PGVS) performs hierarchical, group-aware sparse selection over exogenous variables using a domain-informed physical prior and sparsemax activations, producing a compact, condition-adaptive exogenous context. Exogenous-Conditioned Regime Refinement (ECRR) routes the forecast through learned regime experts that apply gain--bias calibration and horizon-specific corrections via a mixture-of-experts paradigm. Experiments on three wind farms spanning different climates, capacities (66--200 MW), and exogenous dimensions (11--13 variables) demonstrate that ECTO achieves the lowest MSE across all sites, with relative improvements over the strongest baseline ranging from 2.2% to 5.2%, widening to 8.6% at the longer prediction horizon ($H=32$). Ablation analysis confirms that each exogenous-related component contributes positively (PGVS +1.84%, ECRR +2.86%), and interpretability analysis reveals that PGVS learns physically meaningful, site-specific variable selection patterns, while ECRR converges to well-separated calibration strategies consistent across sites.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper proposes ECTO, a framework for ultra-short-term wind power forecasting that decomposes exogenous variable modeling into Physically-Grounded Variable Selection (PGVS) using a domain-informed physical prior with hierarchical group-aware sparse selection via sparsemax, and Exogenous-Conditioned Regime Refinement (ECRR) employing a mixture-of-experts paradigm for gain-bias calibration and horizon-specific corrections. Experiments across three wind farms (66-200 MW, 11-13 exogenous variables, different climates) report that ECTO attains the lowest MSE, with relative gains of 2.2-5.2% over the strongest baseline (widening to 8.6% at H=32), supported by ablations showing +1.84% from PGVS and +2.86% from ECRR, plus interpretability results indicating physically meaningful selections.

Significance. If the physical prior in PGVS proves stable, the method offers a principled way to inject domain knowledge into variable selection for non-stationary forecasting tasks, potentially benefiting grid operations. The combination of sparsemax-based selection and regime experts addresses condition-dependent exogenous effects more explicitly than uniform mixing or PCA. The modest but consistent gains and positive ablations are encouraging, though limited site diversity tempers broader claims of robustness.

major comments (1)
  1. Experiments section: All results are confined to the same three wind farms with no leave-one-farm-out, no additional unseen sites, and no explicit transfer or cross-climate tests. This directly undermines the central claim that the domain-informed physical prior in PGVS yields a selection pattern that remains useful and stable across operating conditions and horizons, as the reported 2.2-5.2% improvements and +1.84% PGVS ablation contribution could arise from site-specific fitting rather than the asserted generalization.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and the opportunity to address concerns regarding the experimental validation of generalization in ECTO. We respond to the major comment below.

read point-by-point responses

  1. Referee: Experiments section: All results are confined to the same three wind farms with no leave-one-farm-out, no additional unseen sites, and no explicit transfer or cross-climate tests. This directly undermines the central claim that the domain-informed physical prior in PGVS yields a selection pattern that remains useful and stable across operating conditions and horizons, as the reported 2.2-5.2% improvements and +1.84% PGVS ablation contribution could arise from site-specific fitting rather than the asserted generalization.

    Authors: We appreciate the referee's point on the scope of validation. The three wind farms were deliberately chosen to span distinct climates, capacities (66-200 MW), and exogenous variable counts (11-13), providing evidence that the physical prior in PGVS produces interpretable, physically consistent selections at each site rather than arbitrary site-specific fits. The prior itself encodes general meteorological relationships (e.g., wind vector components and stability indicators) that are not tuned to individual locations, and the hierarchical sparsemax selection allows condition-adaptive yet constrained choices. Ablations show the PGVS contribution is positive and consistent across all three sites. Nevertheless, we agree that explicit leave-one-farm-out or transfer experiments on additional unseen sites would offer stronger support for cross-condition stability. In the revised manuscript we will (i) expand the discussion of how the domain-informed prior promotes generalization beyond the reported sites and (ii) add an explicit limitations paragraph acknowledging the current site diversity and the value of broader cross-climate testing in future work. These changes will temper the generalization language while preserving the empirical findings. revision: partial

Circularity Check

0 steps flagged

No significant circularity detected in derivation or claims

full rationale

The paper's central results consist of empirical MSE improvements on held-out test data from three wind farms, obtained by comparing the proposed ECTO model (with PGVS and ECRR modules) against external baselines. PGVS incorporates a domain-informed physical prior for variable selection, but this prior is presented as an external input rather than being fitted or derived from the target performance metric itself. No equations or steps reduce a reported prediction to a fitted parameter by algebraic construction, and no load-bearing premise collapses to a self-citation chain. The reported gains (2.2-5.2%, widening at H=32) and ablations are therefore not forced by the model's own inputs or definitions; they remain falsifiable against independent data and baselines. This is the expected non-finding for a standard empirical ML architecture paper.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The framework relies on standard neural-network training assumptions plus an unstated physical prior for variable grouping; no new physical constants or entities are introduced.

free parameters (2)
  • sparsemax temperature / temperature scaling
    Controls sparsity level in PGVS; must be chosen or tuned and directly affects which exogenous variables survive selection.
  • number of regime experts and gating temperature
    Determines how many calibration strategies ECRR learns and how sharply it switches between them.
axioms (1)
  • domain assumption A domain-informed physical prior exists that correctly groups meteorological variables for sparse selection across sites and conditions.
    Invoked to justify the hierarchical group-aware selection inside PGVS.

pith-pipeline@v0.9.0 · 5837 in / 1263 out tokens · 37014 ms · 2026-05-21T08:19:29.054217+00:00 · methodology

Review history (2 revisions) →

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

38 extracted references · 38 canonical work pages

  1. [1]

    R. Tawn, J. Browell, A review of very short-term wind and solar power forecasting, Renewable and Sustainable Energy Reviews 153 (2022) 111758

    work page 2022

  2. [2]

    R. A. Jacobs, M. I. Jordan, S. J. Nowlan, G. E. Hinton, Adaptive mix- tures of local experts, Neural Computation 3 (1) (1991) 79–87

    work page 1991

  3. [3]

    Y. Wang, R. Zou, F. Liu, L. Zhang, Q. Liu, A review of wind speed and wind power forecasting with deep neural networks, Applied Energy 304 (2021) 117766

    work page 2021

  4. [4]

    H. Liu, C. Chen, Data processing strategies in wind energy forecasting models and applications: A comprehensive review, Applied Energy 249 (2019) 392–408

    work page 2019

  5. [5]

    Gallego-Castillo, A

    C. Gallego-Castillo, A. Cuerva-Tejero, O. Lopez-Garcia, A review on the recent history of wind power ramp forecasting, Renewable and Sus- tainable Energy Reviews 52 (2015) 1148–1157

    work page 2015

  6. [6]

    Dalton, B

    A. Dalton, B. Bekker, Exogenous atmospheric variables as wind speed predictors in machine learning, Applied Energy 319 (2022) 119257. 41

    work page 2022

  7. [7]

    Kirchner-Bossi, G

    N. Kirchner-Bossi, G. Kathari, F. Porté-Agel, A hybrid physics-based and data-driven model for intra-day and day-ahead wind power forecast- ing considering a drastically expanded predictor search space, Applied Energy 367 (2024) 123375

    work page 2024

  8. [8]

    Gallego-Castillo, E

    C. Gallego-Castillo, E. García-Bustamante, A. Cuerva, J. Navarro, Iden- tifying wind power ramp causes from multivariate datasets: a method- ological proposal and its application to reanalysis data, IET Renewable PowerGeneration9(8)(2015)867–875.DOI:10.1049/iet-rpg.2014.0457

    work page doi:10.1049/iet-rpg.2014.0457 2015

  9. [9]

    M.Wanek, VariablerenewableenergyforecastinginGermany: Reassess- ing simplicity with Bayesian-optimised multilayer perceptrons, Renew- able Energy 262 (2026) 125409

    work page 2026

  10. [10]

    H. Wang, D. Guo, L. Wang, T. Zhou, C. Jia, Y. Liu, A novel frequency sparsedownsamplinginteractiontransformerforwindpowerforecasting, Energy 326 (2025) 136199

    work page 2025

  11. [11]

    A. Zeng, M. Chen, L. Zhang, Q. Xu, Are transformers effective for time series forecasting?, in: Proceedings of the AAAI Conference on Artificial Intelligence, AAAI, 2023, pp. 11121–11128

    work page 2023

  12. [12]

    Y. Nie, N. H. Nguyen, P. Sinha, A. Ravichander, K. Gao, A time se- ries is worth 64 words: Long-term forecasting with transformers, in: International Conference on Learning Representations, ICLR, 2023

    work page 2023

  13. [13]

    Y. Liu, T. Hu, H. Zhang, H. Wu, S. Wang, L. Ma, M. Long, iTrans- former: Inverted transformers are effective for time series forecasting, in: International Conference on Learning Representations, ICLR, 2024

    work page 2024

  14. [14]

    Y. Wang, H. Wu, J. Dong, et al., TimeXer: Empowering transformers for time series forecasting with exogenous variables, in: Advances in Neural Information Processing Systems, NeurIPS, 2024

    work page 2024

  15. [15]

    X. Chen, H. Jin, Y. Huang, Z. Feng, XLinear: A lightweight and ac- curate MLP-based model for long-term time series forecasting with ex- ogenous inputs, in: Proceedings of the AAAI Conference on Artificial Intelligence, AAAI, 2026. DOI: 10.48448/tj9k-jj18

    work page doi:10.48448/tj9k-jj18 2026

  16. [16]

    P. Zhou, Y. Liu, J. Liang, Q. Song, X. Li, CrossLinear: Plug-and-play cross-correlation embedding for time series forecasting with exogenous 42 variables, in: Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD, 2025, pp. 4120–4131. DOI: 10.1145/3711896.3736899

    work page doi:10.1145/3711896.3736899 2025

  17. [17]

    B. Lim, S. O. Arik, N. Loeff, T. Pfister, Temporal fusion transformers for interpretable multi-horizon time series forecasting, International Journal of Forecasting 37 (4) (2021) 1748–1764

    work page 2021

  18. [18]

    Tayal, A

    K. Tayal, A. Renganathan, X. Jia, V. Kumar, D. Lu, ExoTST: Exogenous-aware temporal sequence transformer for time series predic- tion, in: IEEE International Conference on Data Mining, ICDM, 2024, pp. 857–862. DOI: 10.1109/ICDM59182.2024.00105

    work page doi:10.1109/icdm59182.2024.00105 2024

  19. [19]

    Z. Li, X. Qiu, Y. Zhu, X. Wu, J. Hu, C. Guo, B. Yang, GCGNet: Graph-consistent generative network for time series forecasting with ex- ogenous variables, in: International Conference on Learning Represen- tations, ICLR, 2026

    work page 2026

  20. [20]

    A. F. T. Martins, R. F. Astudillo, From softmax to sparsemax: A sparse model of attention and multi-label classification, in: International Con- ference on Machine Learning, ICML, 2016, pp. 1614–1623

    work page 2016

  21. [21]

    J. B. Olson, J. S. Kenyon, I. Djalalova, et al., Improving wind energy forecasting through numerical weather prediction model development, Bulletin of the American Meteorological Society 100 (11) (2019) 2201– 2220

    work page 2019

  22. [22]

    Al-Yahyai, Y

    S. Al-Yahyai, Y. Charabi, A. Gastli, Review of the use of numerical weather prediction (NWP) models for wind energy assessment, Renew- able and Sustainable Energy Reviews 14 (9) (2010) 3192–3198

    work page 2010

  23. [23]

    J. Jung, R. P. Broadwater, Current status and future advances for wind speed and power forecasting, Renewable and Sustainable Energy Re- views 31 (2014) 762–777

    work page 2014

  24. [24]

    Gneiting, K

    T. Gneiting, K. Larson, K. Westrick, M. G. Genton, E. Aldrich, Cal- ibrated probabilistic forecasting at the Stateline Wind Energy Center: The regime-switching space-time method, Journal of the American Sta- tistical Association 101 (475) (2006) 968–979. 43

    work page 2006

  25. [25]

    Browell, D

    J. Browell, D. R. Drew, K. Philippopoulos, Improved very short-term spatio-temporal wind forecasting using atmospheric regimes, Wind En- ergy 21 (11) (2018) 968–979

    work page 2018

  26. [26]

    Aziz Ezzat, M

    A. Aziz Ezzat, M. Jun, Y. Ding, Spatio-temporal short-term wind fore- cast: A calibrated regime-switching method, Annals of Applied Statis- tics 13 (3) (2019) 1484–1510

    work page 2019

  27. [27]

    Zhang, Y

    Y. Zhang, Y. Li, G. Zhang, Short-term wind power forecasting approach based on Seq2Seq model using NWP data, Energy 213 (2020) 118371

    work page 2020

  28. [28]

    K. Wang, X. Qi, H. Liu, J. Song, Deep belief network based k-means cluster approach for short-term wind power forecasting, Energy 165, Part A (2018) 840–852

    work page 2018

  29. [29]

    Jiang, Q

    Z. Jiang, Q. Tan, N. Li, J. Che, X. Tan, A novel BiGRU multi-step wind power forecasting approach based on multi-label integration random for- est feature selection and neural network clustering, Energy Conversion and Management 319 (2024) 118904

    work page 2024

  30. [30]

    T. Kim, J. Kim, Y. Tae, C. Park, J.-H. Choi, J. Choo, Reversible in- stance normalization for accurate time-series forecasting against distri- bution shift, in: International Conference on Learning Representations, ICLR, 2022

    work page 2022

  31. [31]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, et al., Attention is all you need, in: Advances in Neural Information Processing Systems, NeurIPS, 2017, pp. 5998–6008

    work page 2017

  32. [32]

    Perez, F

    E. Perez, F. Strub, H. De Vries, V. Dumoulin, A. Courville, FiLM: Visual reasoning with a general conditioning layer, in: Proceedings of the AAAI Conference on Artificial Intelligence, AAAI, 2018

    work page 2018

  33. [33]

    Y. Chen, J. Xu, Solar and wind power data from the Chinese State Grid Renewable Energy Generation Forecasting Competition, Scientific Data 9 (2022) 577. DOI: 10.1038/s41597-022-01696-6

    work page doi:10.1038/s41597-022-01696-6 2022

  34. [34]

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, W. Zhang, In- former: Beyond efficient transformer for long sequence time-series fore- casting, in: Proceedings of the AAAI Conference on Artificial Intelli- gence, AAAI, 2021, pp. 11106–11115. 44

    work page 2021

  35. [35]

    H. Wu, T. Hu, Y. Liu, H. Zhou, J. Wang, M. Long, TimesNet: Temporal 2D-variation modeling for general time series analysis, in: International Conference on Learning Representations, ICLR, 2023

    work page 2023

  36. [36]

    S. Wang, H. Wu, X. Shi, T. Hu, H. Luo, L. Ma, J. Y. Zhang, J. Zhou, TimeMixer: Decomposable multiscale mixing for time series forecasting, in: International Conference on Learning Representations, ICLR, 2024

    work page 2024

  37. [37]

    Hochreiter, J

    S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Com- putation 9 (8) (1997) 1735–1780

    work page 1997

  38. [38]

    K. Cho, B. van Merriënboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, Y. Bengio, Learning phrase representations using RNN encoder–decoder for statistical machine translation, in: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Pro- cessing (EMNLP), 2014, pp. 1724–1734. 45 Appendix. Supplementary material Figure 11: Au...

    work page 2014