pith. sign in

arxiv: 2604.18492 · v1 · submitted 2026-04-20 · 💻 cs.LG · cs.SY· eess.SY

Barrier-enforced multi-objective optimization for direct point and sharp interval forecasting

Worachit Amnuaypongsa , Yotsapat Suparanonrat , Pana Wanitchollakit , Jitkomut Songsiri This is my paper

Pith reviewed 2026-05-10 05:33 UTC · model grok-4.3

classification 💻 cs.LG cs.SYeess.SY
keywords probabilistic forecastingprediction intervalsmulti-objective optimizationlog-barrier lossneural networksnon-crossing intervalssolar irradiance
0
0 comments X

The pith

A neural network can output point forecasts and non-crossing intervals that hit exact target coverage while staying as narrow as possible.

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

The paper shows how to train one neural network model to produce both point predictions and prediction intervals in a single forward pass. It reframes the usual trade-off between coverage and sharpness as a multi-objective problem solved by multi-gradient descent, so no manual loss weights need tuning. A new loss term built from an extended log-barrier function with an adaptive parameter pushes the intervals to meet the required coverage rate without widening them unnecessarily. The method is demonstrated on intra-day solar irradiance data, where it meets the target coverage with narrower intervals than standard losses and remains competitive with larger architectures.

Core claim

Treating point forecasting and interval forecasting as simultaneous objectives, solved by multi-gradient descent together with an extended log-barrier loss that has an adaptive hyperparameter, produces a hybrid network whose outputs are guaranteed to be non-crossing intervals that achieve the target PICP while minimizing interval width; the same training procedure removes the need for post-hoc adjustments or trial-and-error weight selection.

What carries the argument

Extended log-barrier loss with adaptive hyperparameter inside a multi-objective optimization solved by multi-gradient descent.

If this is right

  • The same loss and training loop can be attached to any existing deep network architecture without changing its internal design.
  • Point forecasts and calibrated intervals are obtained from one model rather than two separate models.
  • No post-training calibration step is required to restore target coverage.
  • The loss is scale-independent, so the same procedure applies across different data units and forecasting horizons.

Where Pith is reading between the lines

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

  • The barrier approach may replace manual regularization schedules in other constrained forecasting or regression settings.
  • Because coverage is enforced inside the optimizer, the method could be combined with foundation-model backbones without extra fine-tuning stages.
  • Direct multi-step output removes the need for recursive multi-step strategies that accumulate interval widening.

Load-bearing premise

Multi-gradient descent will reliably locate loss weights that enforce the coverage constraint without forcing the intervals to widen.

What would settle it

Run the method on a fresh dataset and check whether the resulting intervals achieve the stated target coverage rate; if coverage falls short or interval widths exceed those of a carefully tuned baseline, the central claim does not hold.

Figures

Figures reproduced from arXiv: 2604.18492 by Jitkomut Songsiri, Pana Wanitchollakit, Worachit Amnuaypongsa, Yotsapat Suparanonrat.

Figure 1
Figure 1. Figure 1: A proposed point and PI forecasting model architecture. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Penalty function components showing the extended log-barrier and the adaptive parameter [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Geometric interpretation of the minimum-norm problem [ [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: MGDA training process: (top) point and PI losses, total loss, and weighted loss; (bottom) gradient norms, inner [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overview of geographical data sources: (a) Geographical distribution of the 104 solar measurement stations [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: A sample time-series of irradiance data. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Overall performance metrics in objective function comparison. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Solar irradiance forecasts (W/sqm) comparing loss functions at 15 minute (top) and 4-hour (bottom) horizons [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Benchmarking model architectures. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Performance metrics comparison for model architecture benchmarking [PITH_FULL_IMAGE:figures/full_fig_p020_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Performance comparison of base models with and without Chronos 2.0 [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Solar irradiance forecasts (W/sqm) comparing model architectures at 15-minute and 4-hour prediction horizons [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Computational complexity [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗
read the original abstract

This paper proposes a multi-step probabilistic forecasting framework using a single neural-network based model to generate simultaneous point and interval forecasts. Our approach ensures non-crossing prediction intervals (PIs) through a model structure design that strictly satisfy a target coverage probability (PICP) while maximizing sharpness. Unlike existing methods that rely on manual weight tuning for scalarized loss functions, we treat point and PI forecasting as a multi-objective optimization problem, utilizing multi-gradient descent to adaptively select optimal weights. Key innovations include a new PI loss function based on an extended log-barrier with an adaptive hyperparameter to guarantee the coverage, a hybrid architecture featuring a shared temporal model with horizon-specific submodels, and a training strategy. The proposed loss is scale-independent and universally applicable; combined with our training algorithm, the framework eliminates trial-and-error hyperparameter tuning for balancing multiple objectives. Validated by an intra-day solar irradiance forecasting application, results demonstrate that our proposed loss consistently outperforms those in current literature by achieving target coverage with the narrowest PI widths. Furthermore, when compared against LSTM encoder-decoder and Transformer architectures--including those augmented with Chronos foundation models--our method remains highly competitive and can be seamlessly adapted to any deep learning structure.

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

2 major / 2 minor

Summary. The manuscript presents a multi-step probabilistic forecasting framework using a single neural network to generate simultaneous point forecasts and non-crossing prediction intervals. It formulates the task as a multi-objective optimization problem solved via multi-gradient descent to adaptively balance objectives without manual weight tuning. Key elements include a hybrid architecture with shared temporal modeling and horizon-specific submodels, a training strategy, and a new PI loss based on an extended log-barrier function with an adaptive hyperparameter intended to strictly enforce target PICP while maximizing sharpness. The loss is claimed to be scale-independent and universally applicable. Validation on intra-day solar irradiance forecasting shows the proposed loss outperforming literature methods by achieving target coverage with the narrowest intervals, while remaining competitive with LSTM encoder-decoder, Transformer, and Chronos-augmented baselines.

Significance. If the adaptive extended log-barrier loss reliably delivers exact target PICP on test data without post-hoc adjustments or sharpness trade-offs, and multi-gradient descent consistently identifies effective objective weights, the work would meaningfully address the common challenge of hyperparameter tuning in interval forecasting methods. The structural enforcement of non-crossing intervals and scale-independent loss could enable broader adoption across forecasting architectures. The approach builds on standard multi-gradient descent with novel loss and architecture elements, but its practical impact hinges on verifying the strict-enforcement claim.

major comments (2)
  1. [Abstract] Abstract: The central claim that the extended log-barrier loss with adaptive hyperparameter 'guarantee[s] the coverage' and enables the framework to 'strictly satisfy a target coverage probability' while 'eliminat[ing] trial-and-error hyperparameter tuning' is load-bearing. Because the barrier term is incorporated into a scalarized objective optimized by multi-gradient descent, the mechanism encourages rather than hard-constrains PICP; any mismatch between the adaptation schedule and training dynamics can produce coverage deviations on unseen data. This creates a correctness risk for the assertion of strict satisfaction without post-training adjustments.
  2. [Abstract] Abstract (validation claims): The assertion that the proposed loss 'consistently outperforms those in current literature by achieving target coverage with the narrowest PI widths' and remains 'highly competitive' against LSTM, Transformer, and Chronos baselines lacks the experimental details needed for evaluation. Specifics on run counts, statistical significance tests, baseline hyperparameter choices, ablation results for the adaptive hyperparameter, and full PICP/MPIW/point-forecast metric tables are required to substantiate outperformance.
minor comments (2)
  1. [Abstract] Abstract: The hybrid architecture and training strategy are described at a high level; explicit details on the output parameterization that structurally enforces non-crossing intervals would aid reproducibility and clarity.
  2. [Abstract] Abstract: Adding citations to prior uses of log-barrier or adaptive-penalty methods in multi-objective neural network training would better situate the novelty of the extended log-barrier loss.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below, providing clarifications and committing to specific revisions that strengthen the manuscript without altering its core contributions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that the extended log-barrier loss with adaptive hyperparameter 'guarantee[s] the coverage' and enables the framework to 'strictly satisfy a target coverage probability' while 'eliminat[ing] trial-and-error hyperparameter tuning' is load-bearing. Because the barrier term is incorporated into a scalarized objective optimized by multi-gradient descent, the mechanism encourages rather than hard-constrains PICP; any mismatch between the adaptation schedule and training dynamics can produce coverage deviations on unseen data. This creates a correctness risk for the assertion of strict satisfaction without post-training adjustments.

    Authors: We appreciate the referee's precise distinction between hard constraints and the barrier formulation. The extended log-barrier is constructed so that the penalty grows without bound as PICP deviates from the target, and the adaptive hyperparameter is updated at each epoch to keep the barrier active and dominant when coverage is at risk. Multi-gradient descent then allocates gradient effort to the barrier objective whenever it is most violated, which in practice produces exact target coverage on the held-out test set without post-hoc recalibration. Nevertheless, we acknowledge that this is an optimization-based enforcement rather than a provably hard constraint for arbitrary data distributions. In the revision we will (i) replace 'guarantee' and 'strictly satisfy' in the abstract with 'effectively enforces' and 'achieves exact target coverage', (ii) add a short subsection deriving the barrier's limiting behavior and the adaptation rule, and (iii) report coverage statistics across multiple random seeds to illustrate robustness to training dynamics. revision: partial

  2. Referee: [Abstract] Abstract (validation claims): The assertion that the proposed loss 'consistently outperforms those in current literature by achieving target coverage with the narrowest PI widths' and remains 'highly competitive' against LSTM, Transformer, and Chronos baselines lacks the experimental details needed for evaluation. Specifics on run counts, statistical significance tests, baseline hyperparameter choices, ablation results for the adaptive hyperparameter, and full PICP/MPIW/point-forecast metric tables are required to substantiate outperformance.

    Authors: We agree that the current experimental reporting is insufficient to support the outperformance claims. The revised manuscript will include: (a) results aggregated over at least five independent runs with different random seeds, (b) paired statistical significance tests (t-tests or Wilcoxon signed-rank) on PICP, MPIW, and point-forecast errors, (c) explicit description of the hyperparameter search procedure used for every baseline (grid ranges or literature defaults), (d) an ablation table isolating the adaptive hyperparameter, and (e) complete side-by-side tables of PICP, MPIW, MAE, and RMSE for all methods on the solar dataset. These additions will be placed in the main experimental section and will directly substantiate the statements in the abstract. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained with external validation

full rationale

The paper introduces a novel extended log-barrier PI loss with adaptive hyperparameter, a hybrid shared-temporal architecture, and multi-gradient descent for weight adaptation as original elements. These are not defined in terms of the target outputs (e.g., no self-definitional reduction where coverage is fitted then renamed as a prediction). Performance claims rest on empirical comparisons to external baselines, LSTM/Transformer models, and Chronos-augmented variants on solar irradiance data, rather than any load-bearing self-citation chain or ansatz smuggled from prior author work. The derivation chain therefore remains independent of its own fitted results.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the new loss function enforcing coverage by construction and on the effectiveness of multi-gradient descent for weight selection; these are not derived from first principles but introduced as design choices.

free parameters (1)
  • adaptive hyperparameter in log-barrier loss
    Adaptive hyperparameter introduced to guarantee target coverage probability; its value is adjusted during training.
axioms (1)
  • domain assumption Multi-gradient descent can adaptively select optimal weights for balancing point forecasting and PI sharpness objectives.
    Invoked to eliminate manual weight tuning in the multi-objective setup.

pith-pipeline@v0.9.0 · 5538 in / 1291 out tokens · 39891 ms · 2026-05-10T05:33:14.568902+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

60 extracted references · 60 canonical work pages

  1. [1]

    2022 30th European Signal Processing Conference (EUSIPCO) , pages=

    Constrained deep networks: Lagrangian optimization via log-barrier extensions , author=. 2022 30th European Signal Processing Conference (EUSIPCO) , pages=. 2022 , organization=

    work page 2022

  2. [2]

    2024 18th International Conference on Probabilistic Methods Applied to Power Systems (PMAPS) , pages=

    Probabilistic Solar Power Forecasting Using Multi-Objective Quantile Regression , author=. 2024 18th International Conference on Probabilistic Methods Applied to Power Systems (PMAPS) , pages=. 2024 , organization=

    work page 2024

  3. [3]

    Energy Conversion and Management: X , volume=

    Neural network-based prediction interval estimation with large width penalization for renewable energy forecasting and system applications , author=. Energy Conversion and Management: X , volume=. 2025 , publisher=

    work page 2025

  4. [4]

    Proceedings of the 35th International Conference on Machine Learning , pages =

    High-Quality Prediction Intervals for Deep Learning: A Distribution-Free, Ensembled Approach , author =. Proceedings of the 35th International Conference on Machine Learning , pages =. 2018 , volume =

    work page 2018

  5. [5]

    Salem and Helge Langseth and Heri Ramampiaro , editor =

    Tárik S. Salem and Helge Langseth and Heri Ramampiaro , editor =. Prediction Intervals: Split Normal Mixture from Quality-Driven Deep Ensembles , volume =. Proceedings of the 36th Conference on Uncertainty in Artificial Intelligence (UAI) , month =

    work page

  6. [6]

    Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics , pages =

    Adaptive, Distribution-Free Prediction Intervals for Deep Networks , author =. Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics , pages =. 2020 , editor =

    work page 2020

  7. [7]

    Exploring uncertainty in regression neural networks for construction of prediction intervals , volume =

    Yuandu Lai and Yucheng Shi and Yahong Han and Yunfeng Shao and Meiyu Qi and Bingshuai Li , doi =. Exploring uncertainty in regression neural networks for construction of prediction intervals , volume =. Neurocomputing , month =

    work page

  8. [8]

    Liu, Siyan and Zhang, Pei and Lu, Dan and Zhang, Guannan , booktitle=

    work page

  9. [9]

    Integrated prediction intervals and specific value predictions for regression problems using neural networks , volume =

    Eli Simhayev and Gilad Katz and Lior Rokach , doi =. Integrated prediction intervals and specific value predictions for regression problems using neural networks , volume =. Knowledge-Based Systems , month =

    work page

  10. [10]

    Deterministic Forecasts and Prediction Intervals for Wind Speed Using Enhanced Multi-Quantile Loss Based Dilated Causal Convolutions , volume =

    Adnan Saeed and Chaoshun Li and Qiannan Zhu and Belal Ahmad , doi =. Deterministic Forecasts and Prediction Intervals for Wind Speed Using Enhanced Multi-Quantile Loss Based Dilated Causal Convolutions , volume =. IEEE Transactions on Sustainable Energy , month =

    work page

  11. [11]

    Enhanced wind speed forecasting for sustainable power systems: A deep learning framework unifying deterministic predictions and uncertainty quantification , volume =

    Adnan Saeed and Chaoshun Li and Saeed Rubaiee and Mohd Danish and Sana Anwar , doi =. Enhanced wind speed forecasting for sustainable power systems: A deep learning framework unifying deterministic predictions and uncertainty quantification , volume =. Energy , month =

    work page

  12. [12]

    Sheppard , doi =

    Giorgio Morales and John W. Sheppard , doi =. Dual Accuracy-Quality-Driven Neural Network for Prediction Interval Generation , year =. IEEE Transactions on Neural Networks and Learning Systems , pages =

    work page

  13. [13]

    Advances in Neural Information Processing Systems , editor =

    Multi-Task Learning as Multi-Objective Optimization , author =. Advances in Neural Information Processing Systems , editor =

    work page

  14. [14]

    Advances in neural information processing systems , volume=

    Pareto multi-task learning , author=. Advances in neural information processing systems , volume=

    work page

  15. [15]

    Proceedings of the 37th International Conference on Machine Learning , pages =

    Multi-Task Learning with User Preferences: Gradient Descent with Controlled Ascent in Pareto Optimization , author =. Proceedings of the 37th International Conference on Machine Learning , pages =. 2020 , editor =

    work page 2020

  16. [16]

    Steepest descent methods for multicriteria optimization , volume =

    Jörg Fliege and Benar Fux Svaiter , doi =. Steepest descent methods for multicriteria optimization , volume =. Mathematical Methods of Operations Research (ZOR) , month =

    work page

  17. [17]

    A Method for Constrained Multiobjective Optimization Based on SQP Techniques , journal =

    Fliege, J\". A Method for Constrained Multiobjective Optimization Based on SQP Techniques , journal =. 2016 , doi =

    work page 2016

  18. [18]

    Revisiting

    Jaggi, Martin , booktitle =. Revisiting. 2013 , editor =

    work page 2013

  19. [19]

    2017 , eprint=

    An Overview of Multi-Task Learning in Deep Neural Networks , author=. 2017 , eprint=

    work page 2017

  20. [20]

    Multiple-gradient descent algorithm

    D. Multiple-gradient descent algorithm. Comptes Rendus. Math

    work page

  21. [21]

    Multiple-Gradient Descent Algorithm

    D. Multiple-Gradient Descent Algorithm. 2009 , MONTH =

    work page 2009

  22. [22]

    Zhang, Qi and Xiao, Peiyao and Zou, Shaofeng and Ji, Kaiyi , booktitle=

    work page

  23. [23]

    Convergence of

    Li, Haochuan and Rakhlin, Alexander and Jadbabaie, Ali , journal=. Convergence of

    work page

  24. [24]

    Winkler , title =

    Robert L. Winkler , title =. Journal of the American Statistical Association , volume =. 1972 , publisher =

    work page 1972

  25. [25]

    and Oumbe, A

    Qu, Z. and Oumbe, A. and Blanc, P. and Espinar, B. and Gesell, G. and Gschwind, B. and Kl. Fast radiative transfer parameterisation for assessing the surface solar irradiance: The. Meteorologische Zeitschrift , volume=. 2017 , publisher=

    work page 2017

  26. [26]

    2025 , eprint=

    Chronos-2: From Univariate to Universal Forecasting , author=. 2025 , eprint=

    work page 2025

  27. [27]

    2023 , institution =

    work page 2023

  28. [28]

    Solar Energy , volume =

    Ineichen, Pierre and Perez, Richard , title =. Solar Energy , volume =. 2002 , publisher =

    work page 2002

  29. [29]

    Deep-learning-based and near real-time solar irradiance map using

    Suwanwimolkul, Suwichaya and Tongamrak, Natanon and Thungka, Nuttamon and Hoonchareon, Naebboon and Songsiri, Jitkomut , journal=. Deep-learning-based and near real-time solar irradiance map using. 2025 , publisher=

    work page 2025

  30. [30]

    GitHub repository , note =

    Improved clear-sky model , year =. GitHub repository , note =

    work page

  31. [31]

    Probabilistic solar forecasting benchmarks on a standardized dataset at

    Yang, Dazhi and van der Meer, Dennis and Munkhammar, Joakim , journal=. Probabilistic solar forecasting benchmarks on a standardized dataset at. 2020 , publisher=

    work page 2020

  32. [32]

    Renewable and Sustainable Energy Reviews , volume=

    The value of solar forecasts and the cost of their errors: A review , author=. Renewable and Sustainable Energy Reviews , volume=. 2024 , publisher=

    work page 2024

  33. [33]

    Advances in Applied Energy , volume=

    Advances in solar forecasting: Computer vision with deep learning , author=. Advances in Applied Energy , volume=. 2023 , publisher=

    work page 2023

  34. [34]

    2021 , publisher=

    Yang, Dazhi and van der Meer, Dennis , journal=. 2021 , publisher=

    work page 2021

  35. [35]

    Operating Reserve Quantification Using Prediction Intervals of Wind Power: An Integrated Probabilistic Forecasting and Decision Methodology , volume =

    Changfei Zhao and Can Wan and Yonghua Song , doi =. Operating Reserve Quantification Using Prediction Intervals of Wind Power: An Integrated Probabilistic Forecasting and Decision Methodology , volume =. IEEE Transactions on Power Systems , month =

    work page

  36. [36]

    2018-2024 , note =

    Vladislav Sovrasov , title =. 2018-2024 , note =

    work page 2018

  37. [37]

    Ke, Guolin and Meng, Qi and Finley, Thomas and Wang, Taifeng and Chen, Wei and Ma, Weidong and Ye, Qiwei and Liu, Tie-Yan , booktitle =. Light

    work page

  38. [38]

    2024 , note =

    Long Short-Term Memory (LSTM) - PyTorch Documentation , author =. 2024 , note =

    work page 2024

  39. [39]

    2024 , note =

    Transformer - PyTorch Documentation , author =. 2024 , note =

    work page 2024

  40. [40]

    arXiv , year =

    Paszke, Adam and Gross, Sam and Massa, Francisco and Lerer, Adam and Bradbury, James and Chanan, Gregory and Killeen, Trevor and Lin, Zeming and Gimelshein, Natalia and Antiga, Luca and Desmaison, Alban and Köpf, Andreas and Yang, Edward and DeVito, Zach and Raison, Martin and Tejani, Alykhan and Chilamkurthy, Sasank and Steiner, Benoit and Fang, Lu and B...

    work page

  41. [41]

    2025 , note =

    Jitkomut Songsiri , title =. 2025 , note =

    work page 2025

  42. [42]

    IEEE Transactions on Neural Networks , month =

    Lower upper bound estimation method for construction of neural network-based prediction intervals , author =. IEEE Transactions on Neural Networks , month =. doi:10.1109/TNN.2010.2096824 , issn =

    work page doi:10.1109/tnn.2010.2096824 2010

  43. [43]

    IEEE Transactions on Power Systems , month =

    An Advanced Approach for Construction of Optimal Wind Power Prediction Intervals , author =. IEEE Transactions on Power Systems , month =. doi:10.1109/TPWRS.2014.2363873 , issn =

    work page doi:10.1109/tpwrs.2014.2363873 2014

  44. [44]

    Energy , month =

    Short-term wind speed interval prediction using improved quality-driven loss based gated multi-scale convolutional sequence model , author =. Energy , month =. doi:10.1016/J.ENERGY.2024.131590 , issn =

    work page doi:10.1016/j.energy.2024.131590 2024

  45. [45]

    2012 , month =

    Etingov, Pavel V and Ma, Jian and Makarov, Yuri V and Subbarao, Krishnappa , title =. 2012 , month =

    work page 2012

  46. [46]

    Cost-Oriented Prediction Intervals: On Bridging the Gap Between Forecasting and Decision , volume =

    Changfei Zhao and Can Wan and Yonghua Song , doi =. Cost-Oriented Prediction Intervals: On Bridging the Gap Between Forecasting and Decision , volume =. IEEE Transactions on Power Systems , month =

    work page

  47. [47]

    1999 , publisher=

    Nonlinear multiobjective optimization , author=. 1999 , publisher=

    work page 1999

  48. [48]

    Gradient-based multi-objective deep learning: Algorithms, theories, applications, and beyond.arXiv preprint arXiv:2501.10945, 2025

    Gradient-based multi-objective deep learning: Algorithms, theories, applications, and beyond , author=. arXiv preprint arXiv:2501.10945 , year=

    work page arXiv

  49. [49]

    Renewable Energy , volume=

    Parametric methods for probabilistic forecasting of solar irradiance , author=. Renewable Energy , volume=. 2018 , publisher=

    work page 2018

  50. [50]

    Solar Energy , volume=

    Probabilistic solar forecasting: Benchmarks, post-processing, verification , author=. Solar Energy , volume=. 2023 , publisher=

    work page 2023

  51. [51]

    IET Renewable Power Generation , volume=

    Point and interval forecasting of solar irradiance with an active Gaussian process , author=. IET Renewable Power Generation , volume=. 2020 , publisher=

    work page 2020

  52. [52]

    Probabilistic solar irradiance forecasting based on

    Li, Xianglong and Ma, Longfei and Chen, Ping and Xu, Hui and Xing, Qijing and Yan, Jiahui and Lu, Siyue and Fan, Haohao and Yang, Lei and Cheng, Yongqiang , journal=. Probabilistic solar irradiance forecasting based on. 2022 , publisher=

    work page 2022

  53. [53]

    Irradiance prediction intervals for

    Scolari, Enrica and Sossan, Fabrizio and Paolone, Mario , journal=. Irradiance prediction intervals for. 2016 , publisher=

    work page 2016

  54. [54]

    Earth Science Informatics , volume=

    A point and interval forecasting of solar irradiance using different decomposition based hybrid models , author=. Earth Science Informatics , volume=. 2023 , publisher=

    work page 2023

  55. [55]

    Energy conversion and management , volume=

    Deterministic and probabilistic forecasting of photovoltaic power based on deep convolutional neural network , author=. Energy conversion and management , volume=. 2017 , publisher=

    work page 2017

  56. [56]

    Energy and AI , pages=

    Interpretable transformer based intra-day solar forecasting with spatiotemporal satellite and numerical weather prediction inputs , author=. Energy and AI , pages=. 2025 , publisher=

    work page 2025

  57. [57]

    Renewable and Sustainable Energy Reviews , volume=

    Multi-step photovoltaic power forecasting using transformer and recurrent neural networks , author=. Renewable and Sustainable Energy Reviews , volume=. 2024 , publisher=

    work page 2024

  58. [58]

    Forty-second International Conference on Machine Learning , year=

    A closer look at transformers for time series forecasting: Understanding why they work and where they struggle , author=. Forty-second International Conference on Machine Learning , year=

    work page

  59. [59]

    Proceedings of the AAAI conference on artificial intelligence , pages=

    Are transformers effective for time series forecasting? , author=. Proceedings of the AAAI conference on artificial intelligence , pages=

    work page

  60. [60]

    Annual review of economics , volume=

    Quantile regression: 40 years on , author=. Annual review of economics , volume=. 2017 , publisher=

    work page 2017