acceptodds
Under review as a conference paper at ICLR 2027

Physics as Representation, Not Regularization: Mechanism and Measurement in Probabilistic Solar Forecasting under Distribution Shift

Abstract

Physics-informed machine learning is motivated by two promises: better predictions in-distribution, and better behaviour outside it. We test both in probabilistic solar forecasting by separating two mechanisms for injecting the same physical information into one shared backbone. Representation exposes physically derived quantities (clear-sky irradiance, solar geometry, air mass) to the predictor as inputs; regularization places physics-based penalties in the training objective, where they shape weights and then vanish at inference. Using 42 inverter systems on five campuses, 19 quantile levels and a parameter-matched physics-free control, we find the two mechanisms are not interchangeable. Representation improves CRPS by 2.21% (95% CI [-4.13, -0.45], five seeds, campus-clustered bootstrap), improves all 42 inverters, cuts implausible night-time generation by a factor of 50, and replicates in a gradient-boosting model of a different family (-1.03%, [-1.49, -0.46]). Regularization alone yields -0.61% ([-2.76, +0.35]) — indistinguishable from zero. Under leave-one-campus-out evaluation, however, transfer is heterogeneous rather than uniform: the representation advantage persists on two campuses, is absent on a third, and reverses on a single-inverter fourth. A single fold — the usual practice — could have supported any of these conclusions. We further show that the large cross-population degradation such a protocol reports (+66% on one fold) is a capacity artefact; measured on identical rows it is at most about 9%. With five campuses the smallest attainable pooled p-value is 0.0625, so we report intervals rather than verdicts. The practical conclusion is that how physics enters a forecaster, and how transfer is measured, change the reported answer more than the physics itself.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
physics-informed machine learning, probabilistic forecasting, distribution shift, solar power forecasting, quantile regression, inductive bias, out-of-distribution generalization, time series

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_TKPUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This is a controlled study on UNISOLAR: 42 inverters on 5 campuses, 1-hour-ahead forecasts with 19 quantiles and one GRU. It compares physics supplied as inputs (clear-sky GHI, zenith and air mass) with physics added as loss penalties. The design includes a parameter-matched control, a LightGBM replication, a campus-clustered bootstrap and leave-one-campus-out folds. Physics inputs help (-2.21% CRPS), and the penalties do not. Transfer varies by campus, and single-fold degradation numbers are mostly a capacity artefact.

Critical strengths

This is the kind of evaluation paper the area needs. The statistics are unusually careful and honest. The paper treats the campus as the unit of inference, and it reports the seed sensitivity, the statistics it discarded, and the fact that LightGBM beats the neural model. It also clearly shows pseudo-replication and the capacity confound in cross-site evaluation. The design is well controlled.

Critical weaknesses

My main issue is with the regularisation arm, because the central claim depends on it. The thresholds were set so that the training targets violate them in only 0.0000% and 0.0244% of rows, and validation pinball loss selects λ = 0.1. In the supplement's Table3_night_violations.csv, the regularised model has about the same share of windows with a negative quantile as the controls (47.5% vs 48.4% and 48.1%). So the regulariser barely binds. The night comparison (L252-255) is not like-for-like either. The night ceiling is 5% of each inverter's 99th-percentile output, while the metric counts medians above 0.1 kWh/15 min. So for larger inverters (campuses 4 and 5, at about 17.8 and 6.1), a night median between these two levels is counted but not penalised. Only λ = 0.1 is reported on test, and the paper does not test stronger λ, hard output constraints or a clear-sky-index target. As it stands, the results show that this loose penalty set is ineffective, not that physics as regularisation is.

The scope is also narrow for the title: one dataset, one climate, one lead time and one small backbone. Campus 1 holds 66% of the test rows, and with five clusters the bootstrap p = 0.015 is not reliable. The paper itself notes that a campus-level test cannot go below 0.0625 (see also Cameron, Gelbach & Miller, 2008). Physics inputs are already standard in PV forecasting, so the positive result is expected. The mechanism and transfer claims are the novel part, and they remain unresolved.

This is a careful paper, and I would like to see it improved. For now, the central claim is not tested well enough, which is the main reason for my score.

Questions for authors

The questions that would most change my view: What were the validation results of the λ search, and what are the test results with λ = 1 and λ = 10, or with a hard constraint? How often is each penalty active, and how does a clear-sky-index arm compare? Does the headline hold with a wild cluster bootstrap or campus-level randomisation? Strong answers here would move me towards acceptance.

Also: does the control receive the campus identity or coordinates? If not, part of the gain may be location information. Does the ranking hold at other lead times?

Reviewer_sYPtOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper considers the physics-informed learning settings where the physics enters as an input representation rather than a regularization in the learning pipeline. In particular, it investigates how different ways to represent the physical information and to measure transferability could change the reported solution more than the presence of physics itself.

Critical strengths

I think the paper idea is interesting and somewhat easy to follow. Understanding different ways to represent the physical information in the learning pipeline and different ways to measure transferability for out-of-distribution generalization is important in physics-informed learning. However, testing only on one one dataset, one university, one climate is very narrow.

Critical weaknesses

Regarding the results, it seems that when the physics is provided through a regularization it doesn’t seem to improve the learning accuracy. The paper also seems to overlook some theoretical results in the literature of physics-informed learning:

  1. A. Scampicchio, LF. Toso, R. Rickenbach, J. Anderson, M. N. Zeilinger, “Physics-informed learning under mixing: How physical knowledge speeds up learning” ICLR 2026
  2. Doumeche, N., F. Bach, G. Biau, and C. Boyer. “Physics-informed machine learning as a kernel method”. COLT 2024.
  3. Diego Marcondes Complexity Dependent Error Rates for Physics-informed Statistical Learning via the Small-ball Method. 2025.

Where they proved that if the physical knowledge is well-aligned with the ground truth model then it should speedup learning. That brings the question of whether the authors attempt to validate whether their physical information is clearly well aligned with the “ground-truth” model and whether there exists other physical lows and information that could provide the improvement that isn’t seen in this validation.

Questions for authors

In Figure 1, I would suggest the authors to provide descriptions for the quantities used since the figure appears in the second page but some of the quantities are only introduced later in the text.

It also seems that more campuses and more training runs would strengthen transfer and comparisons. Perhaps the authors can provide more validation for the refined version of the paper.

It is not super clear to me why the simple concatenation transfers better than gating, perhaps the authors could provide more intuitions in the refined version of the paper.

The paper seems to be missing the position with respect to a broader literature of physics-informed learning and regularized empirical risk minimization in learning theory to also support the empirical findings of the paper.

Reviewer_cxWuOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies two ways of incorporating physics into probabilistic solar forecasting: providing physical quantities as input features and imposing physics-based penalties during training. Using 42 inverter systems across five campuses, the authors compare these approaches under a shared GRU backbone and training budget, with additional LightGBM experiments.

Critical strengths
  1. Clear and relevant experimental question. Separating physics-based features from loss penalties, while controlling architecture and training budget, provides more informative evidence than evaluating a combined “physics-informed” model.
  2. Valuable evaluation analysis. Same-row comparisons reduce confounding from inverter capacity and test-population differences, while evaluating all five campuses exposes the limitations of single-fold conclusions.
  3. Transparent uncertainty reporting. The authors account for within-campus dependence and openly report seed sensitivity, unsuccessful transfer, calibration limitations, and stronger LightGBM performance.
  4. Cross-model validation. The improvement from physical features also appears in LightGBM, suggesting that the benefit is not specific to the GRU architecture.
Critical weaknesses
  1. The central mechanism is not fully isolated. Continuous physical input features and output penalties encode different information and inductive biases. The results establish the relative effectiveness of the tested implementations, rather than a general superiority of representation over regularization. Stronger penalty baselines and a combined feature-plus-regularization arm are needed.
  2. Some conclusions overinterpret statistical non-significance. A confidence interval containing zero does not establish that regularization is ineffective or that gated representation loses its transfer benefit. A direct paired comparison between representation and regularization would better support the main claim.
  3. Cross-campus evidence remains limited. Five highly imbalanced campuses provide few independent units, and overlapping training sets introduce dependence across LOCO folds. Additional seeds cannot compensate for limited site diversity; campus-balanced results and sensitivity analyses would strengthen the conclusions.
  4. The transfer protocol contains target-domain information. Regularization hyperparameters were selected using historical validation data from all campuses, and held-out-site normalization uses historical generation from the target campus. These assumptions should be distinguished from strict unseen-site deployment.
Questions for authors

See critical weaknesses

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.