acceptodds
Under review as a conference paper at ICLR 2027

PVCast: Cloud Evolution for Joint Probabilistic Forecasting of Distributed Solar Power

Abstract

Evolving clouds are an important source of PV power variability, and uncertainty in their evolution affects forecasts across locations and time. These effects shape both marginal power distributions and their spatiotemporal dependence. Prior work largely treats uncertainty in cloud evolution and dependence among PV outputs as separate modeling problems. We introduce PVCast, a joint probabilistic framework that couples marginal PV forecasts and spatiotemporal dependence through shared cloud evolution. PVCast models this evolution with learned transport-source-sink dynamics and represents its uncertainty with an ensemble of future cloud-field sequences. Within each sequence, local cloud conditions inform marginal power forecasts, while the common evolution induces dependence across systems and forecast steps. Across three real-world PV datasets, PVCast outperforms strong time-series, multimodal, and joint probabilistic baselines in marginal and aggregate forecasting, while better capturing the spatial and temporal dependence of joint power trajectories.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Solar forecast, Time series forecast

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_Hnj9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PVCast forecasts the joint distribution of power from multiple PV systems. It generates possible cloud trajectories with transport and source-sink dynamics, and uses them to predict system-level power quantiles and couple the samples across sites and time. Experiments on UKPV, MMSP and PVOD evaluate individual power forecasts, aggregate power and sequences of power changes.

Critical strengths
  1. The use of shared cloud trajectories makes sense for this problem: a moving cloud can affect several PV systems at different times. I particularly like that the sampling construction in Section 3.4 couples the outputs without changing their conditional marginal distributions. This also makes the dependence ablation in Table 3 easier to interpret.

  2. The aggregate forecasting results are encouraging. In Table 1, UKPV CRPS-sum falls from 0.0165 for MTGNN to 0.0119 for PVCast, and PVCast also has the lowest CRPS-sum on MMSP and PVOD. In Table 2, the full cloud-evolution model performs better on aggregate forecasting than either the advection-only or source-sink-only variant.

Critical weaknesses
  1. My main concern is what Table 3 establishes about cloud conditioning. Removing spatial or temporal probability-level coupling makes the corresponding scores worse, but this does not show whether the loadings need to depend on the predicted clouds. In Appendix E.1, even independent sampling retains the shared cloud trajectories and conditional quantiles, so it does not remove all cloud-induced dependence. The no-cloud experiment in Table 2 also changes the marginal forecasts. I would like to see an ablation that keeps the cloud trajectories and conditional quantiles fixed, but replaces the cloud-conditioned loadings with learned, cloud-independent ones.

  2. Some details needed to reproduce the second training stage are missing. Section 3.5 gives three loss terms for fitting the dependence parameters, but I could not find their weights or how those weights were selected. Table 4 specifies the number of cloud trajectories S, but not the number of conditional samples R per trajectory.

Questions for authors
  1. How does PVCast compare with learned, cloud-independent spatial loadings when the cloud trajectories, conditional quantiles, fitting objective, sampling budget and temporal correlation settings are kept the same?

  2. Could you give the three second-stage loss weights, their selection procedure and R? I read the two-stage formulation as keeping the cloud and quantile model fixed while fitting dependence. Please confirm this and clarify which split is used to fit those parameters.

Reviewer_Y8n5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PVCast for joint probabilistic forecasting of distributed solar power, which uses uncertain cloud trajectories to condition both power distributions and their spatial and temporal dependence. Experiments demonstrate the effectiveness of PVCast.

Critical strengths

1.Using shared uncertain cloud evolution to connect marginal power distributions and their dependence is interesting.

2.The experiments are presented in considerable detail.

Critical weaknesses
  1. The dependence ablations are unclear. Removing shared factors or temporal correlations may not ensure independence when systems share stochastic cloud trajectories. The authors should clarify the independence assumptions and sampling strategy.

  2. Table 3 lacks a cloud-independent factor baseline with matched marginals, factor rank, and calibration settings to isolate the contribution of cloud-conditioned dependence modeling.

Questions for authors
  1. How sensitive is PVCast's joint forecasting performance to errors in the predicted cloud trajectories, particularly under rapidly changing cloud conditions?
1 comment

Thank you for the detailed responses and additional experiments. The clarified sampling controls and matched cloud-independent baseline have alleviated some of my concerns and strengthened the empirical evidence. However, I still consider the methodological contribution incremental, primarily extending established dependence modeling through cloud conditioning. The additional results have not sufficiently changed my assessment of its novelty. I therefore maintain my original score.

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

PVCast forecasts the joint distribution of power from distributed PV systems. It estimates cloud motion from satellite history and uses semi-Lagrangian transport to move existing cloud structures. Source and sink fields model local cloud formation and dissipation. Perturbing the motion and source-sink parameters generates future cloud trajectories. For each trajectory, the shared power decoder predicts conditional quantile functions. The cloud-conditioned spatiotemporal dependence model generates coupled probability levels. The quantile functions are evaluated at these levels to sample joint power trajectories. The decoder also uses historical power, NWP and system information. Training first fits cloud evolution and conditional power distributions, then fits dependence parameters. Experiments cover UKPV, MMSP and PVOD, with 841, 88 and 10 systems. The reported gains in CRPS-sum and ES are larger than those in marginal CRPS.

Critical strengths

The consistent CRPS-sum gains across the three datasets. They are larger than the gains in marginal CRPS, suggesting that the method is particularly useful for aggregate power forecasting. The UKPV experiment tests this on 841 systems, where dependence across systems matters for the uncertainty in total generation.

Critical weaknesses
  1. My main concern is whether cloud-conditioned spatiotemporal dependence outperforms a conventional dependence model when the conditional quantile functions are kept fixed. Table 3 removes shared cross-system factors or sets the AR(1) coefficients to zero. These tests show that dependence modeling helps, but do not isolate the benefit of cloud conditioning. The cloud-evolution ablations in Table 2 also change the cloud inputs to the power decoder. A comparison with a dependence model fitted without cloud inputs is needed to establish the benefit of cloud conditioning.

  2. The cloud ensemble also needs quantitative verification. Cloud-field skill in Equations 22 and 23 uses the unperturbed forecast, while Figure 5 illustrates only one event. This leaves little evidence about the ensemble spread and spatiotemporal dependence across events. The aggregate and ramp scores measure the final power forecasts, to which cloud sampling, conditional power distributions and the fitted dependence model all contribute. A comparison using the unperturbed cloud forecast would help identify the gain from sampling cloud trajectories.

Questions for authors
  1. How does cloud-conditioned dependence compare with a static copula or a copula fitted to historical forecast errors, using the same conditional quantile functions? Please match the number of shared Gaussian factors, the data used to fit dependence parameters and the number of sampled joint power trajectories.

  2. Can you evaluate the cloud ensemble against observed cloud fields and compare power forecasts obtained with sampled cloud trajectories against those obtained with the unperturbed cloud forecast? The latter comparison should use the same power-decoder architecture, dependence model and total number of sampled joint power trajectories.

1 comment

Thanks for the responses which address some of my concerns. I appreciate the efforts. However, I still have concerns that the core ideas do not show enough novelty. In additon, the amount of revisions is major and there seems no mechanism that we can evaluate the final form of the paper. Thus, I'll keep my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.