pith. sign in

arxiv: 2604.16614 · v1 · submitted 2026-04-17 · 📡 eess.SY · cs.SY

CVaR-Guided Decision-Focused Learning and Risk-Triggered Re-Optimization for Two-Stage Robust Microgrid Operation

Tingwei Cao , Yan Xu This is my paper

Pith reviewed 2026-05-10 07:25 UTC · model grok-4.3

classification 📡 eess.SY cs.SY
keywords microgrid operationtwo-stage robust optimizationdecision-focused learningCVaRprobabilistic forecastingload uncertaintyrisk-triggered re-optimizationKKT implicit differentiation
0
0 comments X

The pith

CVaR-guided decision-focused learning aligns probabilistic forecasts with two-stage robust microgrid scheduling while cutting online computation.

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

The paper establishes that a forecasting model trained to emphasize tail-risk samples through CVaR, then refined by operational feedback from a differentiable surrogate of the two-stage robust optimization, produces uncertainty sets that improve the quality of downstream robust schedules. A sympathetic reader would care because standard predict-then-optimize pipelines frequently produce forecasts whose quality does not translate into better decisions under load uncertainty, leaving microgrid operators with higher costs or unmitigated tail risks. The framework adds a risk-triggered re-optimization step that re-solves the remaining-horizon problem only when schedule mismatch grows large, preserving most of the benefit of continuous re-optimization at far lower daily effort. If these mechanisms work as described, operators obtain forecasts and schedules that jointly reduce operating cost and tail exposure without requiring repeated full solves at every time step.

Core claim

The central claim is that converting multi-quantile forecasts into load uncertainty sets, training the forecaster with a CVaR objective on difficult samples, and closing the forecast-decision loop through a convex regularized surrogate TSRO model plus smooth regret loss (with KKT-based implicit differentiation) yields better probabilistic accuracy, lower operating costs, and stronger tail-risk control than benchmark methods on modified IEEE 33-bus and 69-bus systems; the risk-triggered re-optimization mechanism then delivers this performance with less than 0.5 percent higher cost and up to 91 percent lower daily solution time.

What carries the argument

The CVaR-guided decision-focused learning loop that uses a convex regularized surrogate of the two-stage robust optimization together with a smooth regret loss to enable KKT implicit differentiation, plus the risk-triggered re-optimization trigger that activates only on significant schedule mismatch.

If this is right

  • Probabilistic forecasts become more reliable precisely on the high-risk operating days that matter most for robust scheduling.
  • The resulting two-stage robust schedules achieve lower expected operating cost and better tail-risk mitigation than forecasts trained without decision feedback.
  • Near-full re-optimization performance is retained while daily solution time falls by up to 91 percent.
  • The forecast-decision gap is narrowed without sacrificing the original robustness properties of the two-stage model.
  • Online deployment becomes practical because re-optimization is invoked selectively rather than at every interval.

Where Pith is reading between the lines

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

  • The same surrogate-plus-CVaR pattern could be tested on robust unit-commitment or transmission-constrained problems where uncertainty sets are also derived from quantiles.
  • Replacing the current risk-trigger threshold with a learned policy might further reduce the small residual cost gap while keeping computation low.
  • The approach indicates that decision-focused pipelines can supplant separate forecasting stages in other sequential robust scheduling settings once a convex surrogate is available.
  • Scaling the KKT differentiation step to networks larger than 69 buses would require checking whether the surrogate remains tractable and whether gradient quality holds.

Load-bearing premise

The convex regularized surrogate TSRO model and smooth regret loss supply a sufficiently accurate and differentiable stand-in for the true two-stage robust optimization so that KKT gradients improve the forecaster without distorting robustness guarantees or biasing tail-risk handling.

What would settle it

A side-by-side run on the same IEEE 69-bus instances in which the proposed method's realized operating cost or CVaR exceeds that of continuous full re-optimization by more than 0.5 percent on average would show the surrogate approximation introduces unacceptable decision bias.

Figures

Figures reproduced from arXiv: 2604.16614 by Tingwei Cao, Yan Xu.

Figure 1
Figure 1. Figure 1: Overall framework of the proposed method. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: illustrates the surrogate TSRO mechanism. Through convex relaxation and quadratic regularization, the original TSRO landscape is transformed into a surrogate landscape with a unique solution [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Online operational workflow of the proposed method. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: IEEE 33-bus and 69-bus microgrid topologies. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: RTRO activation and computational performance of different methods. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: Probabilistic load forecasting results on representative typical and [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: 24 h dispatch results of the proposed method on the representative [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
read the original abstract

Microgrid operation is highly vulnerable to short-term load uncertainty, while conventional predict-then-optimize pipelines cannot fully align probabilistic forecasting quality with downstream robust scheduling performance. This paper proposes a CVaR-guided decision-focused learning and risk-triggered re-optimization framework for two-stage robust microgrid operation. A probabilistic load forecasting model first generates multi-quantile outputs, which are converted into prediction intervals to parameterize the load uncertainty set of the downstream two-stage robust optimization (TSRO) model. To improve forecasting reliability under difficult and high-risk operating conditions, a CVaR-guided forecasting objective is introduced to emphasize tail-sensitive samples. To further close the forecast-decision gap, a convex regularized surrogate TSRO model and a smooth regret loss are developed, enabling downstream operational feedback to be propagated to the forecasting model through KKT-based implicit differentiation. For online deployment, a risk-triggered re-optimization mechanism selectively re-solves the remaining-horizon TSRO only when the schedule mismatch becomes significant, avoiding unnecessary online computation. Case studies on modified IEEE 33-bus and 69-bus microgrids demonstrate superior probabilistic forecasting accuracy, operational economy, and tail-risk mitigation over benchmark methods, while preserving near-full-re-optimization performance with less than 0.5% higher operating cost and up to 91% lower daily solution time.

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

3 major / 2 minor

Summary. The paper proposes a CVaR-guided decision-focused learning and risk-triggered re-optimization framework for two-stage robust microgrid operation. A multi-quantile probabilistic forecaster parameterizes the uncertainty set of a TSRO model; a convex regularized surrogate TSRO plus smooth regret loss enables KKT-based implicit differentiation to align forecasts with operational regret; and a risk-triggered mechanism selectively re-solves the remaining-horizon TSRO. Case studies on modified IEEE 33-bus and 69-bus systems report superior probabilistic accuracy, lower operating costs, and tail-risk mitigation versus benchmarks while incurring <0.5% extra cost and up to 91% lower daily solve time relative to full re-optimization.

Significance. If the surrogate approximation is shown to preserve the original robustness guarantees and CVaR tail emphasis, the framework would meaningfully advance decision-focused learning for robust energy-system optimization by directly optimizing downstream performance rather than forecast accuracy alone. The risk-triggered re-optimization offers a practical route to computational savings without sacrificing near-optimal schedules. The use of standard IEEE test systems and explicit quantitative claims on both accuracy and runtime are strengths that facilitate reproducibility.

major comments (3)
  1. [§3.2] §3.2 (Convex Regularized Surrogate TSRO): The claim that the regularized surrogate plus smooth regret loss yields gradients that faithfully optimize the original two-stage robust problem is load-bearing for all performance claims. No error bounds, sensitivity analysis, or comparison of worst-case uncertainty realizations (or CVaR-weighted tail samples) between the surrogate and the true min-max TSRO are provided; without this, it is unclear whether the learned forecasts actually improve the deployed robust schedule or merely the proxy.
  2. [§4] §4 (Case Studies): The reported gains in forecasting accuracy, operating cost, and tail-risk mitigation rest on experimental results that omit data-split details, number of independent runs, statistical significance tests, error bars, and the procedure used to select the CVaR level, regularization weight, and risk-trigger threshold. These omissions make it impossible to assess whether the <0.5% cost gap and 91% time reduction are robust or sensitive to hyperparameter choices and data partitioning.
  3. [§3.3] §3.3 (Smooth Regret Loss): Because the regret loss is computed from the surrogate optimization outcome and the surrogate parameters are tuned on the same operational data used for training, the forecasting improvements may partly reflect in-sample fitting to the downstream objective. This circularity directly threatens the tail-risk mitigation claims and requires either hold-out validation or an explicit analysis of generalization to unseen uncertainty realizations.
minor comments (2)
  1. [Abstract and §4] The abstract and §4 should explicitly name the benchmark methods (e.g., which forecasting models and TSRO solvers) and the precise definition of the prediction-interval-to-uncertainty-set mapping.
  2. [§3.1] Notation for the multi-quantile outputs, prediction intervals, and the CVaR-guided objective could be consolidated into a single equation block for clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment point by point below, providing clarifications and committing to revisions that strengthen the manuscript without overstating current results.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (Convex Regularized Surrogate TSRO): The claim that the regularized surrogate plus smooth regret loss yields gradients that faithfully optimize the original two-stage robust problem is load-bearing for all performance claims. No error bounds, sensitivity analysis, or comparison of worst-case uncertainty realizations (or CVaR-weighted tail samples) between the surrogate and the true min-max TSRO are provided; without this, it is unclear whether the learned forecasts actually improve the deployed robust schedule or merely the proxy.

    Authors: We agree that the absence of explicit error bounds or sensitivity analysis leaves the approximation quality insufficiently justified. The convex regularized surrogate is constructed to retain the min-max structure and enable KKT-based differentiation, but we acknowledge that empirical validation alone is not fully convincing. In the revision we will add a dedicated subsection in §3.2 containing (i) a numerical sensitivity study comparing worst-case uncertainty realizations and objective values between the surrogate and the exact TSRO on the IEEE 33-bus and 69-bus instances, and (ii) a comparison of CVaR-weighted tail samples. These additions will quantify the approximation gap and demonstrate that the learned forecasts translate into improved schedules under the true robust model. revision: yes

  2. Referee: [§4] §4 (Case Studies): The reported gains in forecasting accuracy, operating cost, and tail-risk mitigation rest on experimental results that omit data-split details, number of independent runs, statistical significance tests, error bars, and the procedure used to select the CVaR level, regularization weight, and risk-trigger threshold. These omissions make it impossible to assess whether the <0.5% cost gap and 91% time reduction are robust or sensitive to hyperparameter choices and data partitioning.

    Authors: We fully concur that these experimental details are required for reproducibility and credibility. The revised Section 4 will explicitly state: the chronological data-split ratios and partitioning method, the number of independent runs with standard deviations and error bars on all reported metrics, the results of statistical significance tests (paired t-tests against benchmarks), and the complete hyperparameter selection procedure (grid-search ranges and cross-validation criterion used to choose the CVaR level, regularization weight, and risk-trigger threshold). These additions will allow readers to evaluate the robustness of the <0.5% cost gap and 91% runtime reduction. revision: yes

  3. Referee: [§3.3] §3.3 (Smooth Regret Loss): Because the regret loss is computed from the surrogate optimization outcome and the surrogate parameters are tuned on the same operational data used for training, the forecasting improvements may partly reflect in-sample fitting to the downstream objective. This circularity directly threatens the tail-risk mitigation claims and requires either hold-out validation or an explicit analysis of generalization to unseen uncertainty realizations.

    Authors: The potential for in-sample bias is a legitimate concern. While the manuscript already uses a held-out test set for final evaluation and a validation set for hyperparameter tuning, we recognize that an explicit generalization study is missing. In the revision we will add an analysis that evaluates the learned forecaster on additional out-of-sample uncertainty realizations drawn from the test distribution, together with a direct comparison of in-sample versus out-of-sample regret and tail-risk metrics. This will demonstrate that the reported tail-risk mitigation generalizes beyond the training data. revision: partial

Circularity Check

0 steps flagged

No significant circularity; derivation uses standard decision-focused learning without reducing claims to self-definition or fitted inputs

full rationale

The paper's chain proceeds from a probabilistic forecaster (multi-quantile outputs) to uncertainty sets for a two-stage robust optimization (TSRO), with a CVaR-guided objective and a convex regularized surrogate plus smooth regret loss to enable KKT-based implicit differentiation for end-to-end training. This is an explicit methodological choice to align forecasting with downstream decisions, not a redefinition where outputs equal inputs by construction. Case-study claims of superior accuracy, economy, and tail-risk mitigation are evaluated against external benchmarks on IEEE test systems, with no load-bearing step that renames a fit as a prediction or relies on self-citation chains for uniqueness. The surrogate approximates the original TSRO for differentiability but does not substitute for it in the final performance metrics.

Axiom & Free-Parameter Ledger

3 free parameters · 2 axioms · 0 invented entities

Only the abstract is available, so the ledger is necessarily incomplete; the framework rests on standard robust-optimization assumptions plus several modeling choices whose independence from the reported gains cannot be verified.

free parameters (3)
  • CVaR confidence level
    Controls emphasis on tail samples in the forecasting objective; value not stated in abstract.
  • Regularization weight in surrogate TSRO
    Balances convexity and fidelity to original robust model; tuning details absent.
  • Risk-trigger threshold
    Determines when re-optimization is invoked; directly affects the reported computation savings.
axioms (2)
  • domain assumption Prediction intervals derived from multi-quantile forecasts can faithfully parameterize the uncertainty set for two-stage robust optimization.
    Central to linking the forecasting model to the downstream TSRO.
  • domain assumption The KKT conditions of the convex surrogate remain a valid and stable basis for implicit differentiation even under load uncertainty.
    Required for propagating operational feedback to the forecaster.

pith-pipeline@v0.9.0 · 5544 in / 1617 out tokens · 60970 ms · 2026-05-10T07:25:42.060147+00:00 · methodology

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

30 extracted references · 30 canonical work pages

  1. [1]

    Smart distribution: Coupled microgrids,

    R. H. Lasseter, “Smart distribution: Coupled microgrids,”Proceedings of the IEEE, vol. 99, no. 6, pp. 1074–1082, 2011

    work page 2011

  2. [2]

    Hierarchical control of droop-controlled ac and dc micro- grids—a general approach toward standardization,

    J. M. Guerrero, J. C. Vasquez, J. Matas, L. G. De Vicu ˜na, and M. Castilla, “Hierarchical control of droop-controlled ac and dc micro- grids—a general approach toward standardization,”IEEE Transactions on industrial electronics, vol. 58, no. 1, pp. 158–172, 2010

    work page 2010

  3. [3]

    Robust optimization of microgrid based on renew- able distributed power generation and load demand uncertainty,

    J. Yang and C. Su, “Robust optimization of microgrid based on renew- able distributed power generation and load demand uncertainty,”Energy, vol. 223, p. 120043, 2021

    work page 2021

  4. [4]

    Day-ahead and intraday multi-time scale microgrid scheduling based on light robustness and mpc,

    Y . He, Z. Li, J. Zhang, G. Shi, and W. Cao, “Day-ahead and intraday multi-time scale microgrid scheduling based on light robustness and mpc,”International Journal of Electrical Power & Energy Systems, vol. 144, p. 108546, 2023

    work page 2023

  5. [5]

    A historical- correlation-driven robust optimization approach for microgrid dispatch,

    H. Qiu, W. Gu, X. Xu, G. Pan, P. Liu, Z. Wu, and L. Wang, “A historical- correlation-driven robust optimization approach for microgrid dispatch,” IEEE Transactions on Smart Grid, vol. 12, no. 2, pp. 1135–1148, 2020

    work page 2020

  6. [6]

    Robust operation of microgrids via two-stage coordinated energy storage and direct load control,

    C. Zhang, Y . Xu, Z. Y . Dong, and J. Ma, “Robust operation of microgrids via two-stage coordinated energy storage and direct load control,”IEEE Transactions on Power Systems, vol. 32, no. 4, pp. 2858–2868, 2016

    work page 2016

  7. [7]

    Robust coordination of distributed generation and price-based demand response in microgrids,

    C. Zhang, Y . Xu, Z. Y . Dong, and K. P. Wong, “Robust coordination of distributed generation and price-based demand response in microgrids,” IEEE Transactions on Smart Grid, vol. 9, no. 5, pp. 4236–4247, 2017

    work page 2017

  8. [8]

    Applications of probabilistic forecasting in smart grids: A review,

    H. Khajeh and H. Laaksonen, “Applications of probabilistic forecasting in smart grids: A review,”Applied Sciences, vol. 12, no. 4, p. 1823, 2022

    work page 2022

  9. [9]

    A more efficient microgrid operation through the integration of probabilistic solar forecasts,

    F. Ramahatana, J. L. G. La Salle, P. Lauret, and M. David, “A more efficient microgrid operation through the integration of probabilistic solar forecasts,”Sustainable Energy, Grids and Networks, vol. 31, p. 100783, 2022

    work page 2022

  10. [10]

    Quantifying the value of probabilistic forecasting for power system operation planning,

    Q. Wang, A. Tuohy, M. Ortega-Vazquez, M. Bello, E. Ela, D. Kirk- Davidoff, W. B. Hobbs, D. J. Ault, and R. Philbrick, “Quantifying the value of probabilistic forecasting for power system operation planning,” Applied Energy, vol. 343, p. 121254, 2023

    work page 2023

  11. [11]

    Optimal scheduling of renewable energy microgrids: A robust multi- objective approach with machine learning-based probabilistic forecast- ing,

    D. Aguilar, J. J. Quinones, L. R. Pineda, J. Ostanek, and L. Castillo, “Optimal scheduling of renewable energy microgrids: A robust multi- objective approach with machine learning-based probabilistic forecast- ing,”Applied Energy, vol. 369, p. 123548, 2024

    work page 2024

  12. [12]

    Data-driven robust optimization scheduling for microgrid day-ahead to intra-day operations based on renewable energy interval prediction,

    M. Yang, J. Wang, Y . Chen, Y . Zeng, and X. Su, “Data-driven robust optimization scheduling for microgrid day-ahead to intra-day operations based on renewable energy interval prediction,”Energy, vol. 313, p. 134058, 2024

    work page 2024

  13. [13]

    Smart “predict, then optimize

    A. N. Elmachtoub and P. Grigas, “Smart “predict, then optimize”,” Management Science, vol. 68, no. 1, pp. 9–26, 2022

    work page 2022

  14. [14]

    Risk-aware objective-based forecasting in inertia management,

    H. Zhang, R. Li, Y . Chen, Z. Chu, M. Sun, and F. Teng, “Risk-aware objective-based forecasting in inertia management,”IEEE Transactions on Power Systems, vol. 39, no. 2, pp. 4612–4623, 2024

    work page 2024

  15. [15]

    Toward value-oriented renewable energy forecasting: An iterative learning approach,

    Y . Zhang, M. Jia, H. Wen, Y . Bian, and Y . Shi, “Toward value-oriented renewable energy forecasting: An iterative learning approach,”IEEE Transactions on Smart Grid, vol. 16, no. 2, pp. 1962–1974, 2025

    work page 1962

  16. [16]

    Cost-oriented prediction intervals: On bridging the gap between forecasting and decision,

    C. Zhao, C. Wan, and Y . Song, “Cost-oriented prediction intervals: On bridging the gap between forecasting and decision,”IEEE Transactions on Power Systems, vol. 37, no. 4, pp. 3048–3062, 2022

    work page 2022

  17. [17]

    Smart predict-then-optimize for two- stage linear programs with side information,

    A. S. Estes and J.-P. P. Richard, “Smart predict-then-optimize for two- stage linear programs with side information,”INFORMS Journal on Optimization, vol. 5, no. 3, pp. 295–320, 2023

    work page 2023

  18. [18]

    Feature-driven economic improvement for network-constrained unit commitment: A closed-loop predict-and-optimize framework,

    X. Chen, Y . Yang, Y . Liu, and L. Wu, “Feature-driven economic improvement for network-constrained unit commitment: A closed-loop predict-and-optimize framework,”IEEE Transactions on Power Systems, vol. 37, no. 4, pp. 3104–3118, 2022

    work page 2022

  19. [19]

    Decision-focused learning for power system decision-making under uncertainty,

    H. Zhang, R. Li, Q. Du, J. Tao, S. Pineda, G. Kariniotakis, S. Camal, C. B. Monroc, M. Sun, C. Wan, W. Xu, and F. Teng, “Decision-focused learning for power system decision-making under uncertainty,”IEEE Transactions on Power Systems, vol. 41, no. 1, pp. 307–323, 2026

    work page 2026

  20. [20]

    Decision- focused learning integrated with data-driven robust optimization for energy management,

    G. Wang, Z. Lin, X. Zhong, H. Ji, P. Li, Y . Chen, and J. Yan, “Decision- focused learning integrated with data-driven robust optimization for energy management,”Applied Energy, vol. 408, p. 127343, 2026

    work page 2026

  21. [21]

    Decision-focused learning enhanced by automated feature engineering for energy storage optimisation,

    N. Alkhulaifi, I. Gokay Dogan, T. R. Cargan, A. L. Bowler, D. Pekaslan, N. J. Watson, and I. Triguero, “Decision-focused learning enhanced by automated feature engineering for energy storage optimisation,”Expert Systems with Applications, vol. 302, p. 130554, 2026

    work page 2026

  22. [22]

    Risk-aware prediction-optimization inte- grated method for day-ahead microgrid operation,

    H. Sheng, Y . Xu, and D. Xie, “Risk-aware prediction-optimization inte- grated method for day-ahead microgrid operation,”IEEE Transactions on Smart Grid, 2026

    work page 2026

  23. [23]

    A novel fitted rolling horizon control approach for real-time policy making in microgrid,

    A. Das and Z. Ni, “A novel fitted rolling horizon control approach for real-time policy making in microgrid,”IEEE Transactions on Smart Grid, vol. 11, no. 4, pp. 3535–3544, 2020

    work page 2020

  24. [24]

    Economic model predictive control for microgrid optimization: A review,

    J. Hu, Y . Shan, Y . Yang, A. Parisio, Y . Li, N. Amjady, S. Islam, K. W. Cheng, J. M. Guerrero, and J. Rodr ´ıguez, “Economic model predictive control for microgrid optimization: A review,”IEEE Transactions on Smart Grid, vol. 15, no. 1, pp. 472–484, 2023

    work page 2023

  25. [25]

    Feature-enhanced deep learning method for electric vehicle charging demand probabilistic forecasting of charging station,

    T. Cao, Y . Xu, G. Liu, S. Tao, W. Tang, and H. Sun, “Feature-enhanced deep learning method for electric vehicle charging demand probabilistic forecasting of charging station,”Applied Energy, vol. 371, p. 123751, 2024

    work page 2024

  26. [26]

    Optimal capacitor placement on radial distribution systems,

    M. E. Baran and F. F. Wu, “Optimal capacitor placement on radial distribution systems,”IEEE Transactions on power Delivery, vol. 4, no. 1, pp. 725–734, 1989

    work page 1989

  27. [27]

    Optimization of conditional value- at-risk,

    R. T. Rockafellar, S. Uryasevet al., “Optimization of conditional value- at-risk,”Journal of risk, vol. 2, pp. 21–42, 2000

    work page 2000

  28. [28]

    Differentiable convex optimization layers,

    A. Agrawal, B. Amos, S. Barratt, S. Boyd, S. Diamond, and J. Z. Kolter, “Differentiable convex optimization layers,”Advances in neural information processing systems, vol. 32, 2019

    work page 2019

  29. [29]

    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

    work page 2017

  30. [30]

    Mqtransformer: Multi-horizon forecasts with context dependent and feedback-aware attention,

    C. Eisenach, Y . Patel, and D. Madeka, “Mqtransformer: Multi-horizon forecasts with context dependent and feedback-aware attention,”arXiv preprint arXiv:2009.14799, 2020

    work page arXiv 2009