acceptodds
Under review as a conference paper at ICLR 2027

SolarSpecNet: Phase-Guided Structure–Intensity Modeling for Short-Term Solar Radiation Field Forecasting

Abstract

Short-term solar-radiation forecasting is essential for photovoltaic integration, power scheduling, and environmental analysis. However, most current methods either focus on station-level time-series prediction or treat gridded radiation fields as generic image sequences. Consequently, they fail to explicitly model two tightly coupled aspects of future radiation-field evolution. 1) Where will regions of high and low radiation emerge? 2) How will the radiation intensity at each location change over time? To address this limitation, we propose SolarSpecNet, which explicitly disentangles these two aspects for specialized modeling while preserving their interaction through spectral-spatial fusion. Specifically, a Structural Evolution Branch exploits Fourier phase information to forecast the spatial evolution of radiation patterns, while a Radiative Energy Branch captures radiation-intensity dynamics from both spatial-domain field values and frequency-domain representations. Then, a Spectral-Spatial Fusion Module integrates these complementary predictions through spectral recombination and spatial refinement to produce the final radiation-field forecast. Experiments on gridded radiation fields show that SolarSpecNet consistently improves multi-step forecasting over representative statistical, recurrent, convolutional, spectral, and foundation time-series baselines. Additional visualizations show that its forecasts evolve smoothly, retain clear day–night boundaries, and preserve the main radiation patterns over time. Codes are attached.

Primary area:
learning on time series and dynamical systems
Keywords:
Short-term solar-radiation forecasting, Gridded radiation fields, Spatio-temporal evolution, Spectral-spatial fusion, SolarSpecNet.

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_4kUiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes SolarSpecNet for short-term solar radiation field forecasting. The Structural Evolution Branch (SEB) predicts spatial patterns using Fourier phase, while the Radiative Energy Branch (REB) models radiation intensity. A Spectral-Spatial Fusion Module (SSF) combines their outputs to generate multi-step forecasts. Experiments on ERA5 radiation fields demonstrate improved accuracy over statistical, deep learning, and foundation model baselines. Ablation studies further evaluate the contributions of individual modules and auxiliary losses.

Critical strengths
  1. The paper addresses an important problem in solar radiation field forecasting, which is valuable for photovoltaic generation and energy management. Modeling both spatial structure and radiation intensity is a meaningful research direction.
  2. The paper is well-organized and easy to follow. The motivation, methodology, and experimental results are clearly presented, with helpful figures illustrating the overall framework.
  3. The proposed method combines phase-guided structural prediction, radiative intensity modeling, and spectral-spatial fusion into a unified framework. The overall design is intuitive and provides a reasonable approach to modeling different aspects of radiation-field dynamics.
Critical weaknesses
  1. SEB uses the latest historical amplitude to reconstruct future radiation structures. However, radiation fields exhibit complex spatial dynamics beyond simple translation. The paper does not sufficiently justify whether fixed historical amplitude can provide reliable structural guidance over a 20-hour forecasting horizon.
  2. The w/o SSF experiment replaces the final prediction with a low-frequency spectral reconstruction rather than directly combining SEB and REB outputs. This introduces additional information loss, making it difficult to isolate the contribution of SSF. More controlled comparisons, such as REB-only or simple fusion, are needed.
  3. The proposed framework closely resembles AlphaPre [1] (CVPR 2025), which also uses phase–amplitude modeling and spectral-spatial fusion for spatio-temporal forecasting. However, the paper does not clearly distinguish its technical contributions from this closely related work or include it as a baseline.
  4. Experiments are limited to a single ERA5 dataset with only 120 highly overlapping test samples. Additional evaluations across different seasons, geographical regions, and weather conditions are needed to demonstrate broader generalizability.

Reference [1] Lin, Kenghong, et al. "AlphaPre: Amplitude-phase disentanglement model for precipitation nowcasting." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025.

Questions for authors

Please refer to weakness

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

This paper aims to predict a 20-hour solar radiation field with the historical 5-hour data as input. In the proposed method, the structural branch predicts the Fourier phase, while the energy branch predicts the complete radiance field and spectral amplitude; a fusion module then combines these with the results of low-frequency reconstruction. During training, the final output energy, amplitude, and phase are all subject to supervision. Experiments show the proposed outperform six baselines such as ARIMA, SimVP, ICNO.

Critical strengths
  1. The research task is very important.
  2. The ablation studies verify the effectiveness of the key elements in the proposed method.
Critical weaknesses
  1. My biggest concern about this work is the novelty. Solar radiation field forecasting is an application of video prediction. The key techniques in the proposed SolarSpecNet, such as the phase branch, amplitude branch, and learned fusion, are very similar to AlphaPre (CVPR 2025). Although AlphaPre is proposed for precipitation nowcasting, the authors should discuss the difference carfully, rather than simply swicth a application domain.

  2. The test samples may only cover about 6 days. The experiments on such samples are not convincing, since lots of overlapping windows cannot be regarded as independent weather events.

  3. Why do the authors include time series methods such as ARIMA, LSTM, and Chronos-2 for comparison under a spatiotemporal predictive setting? Is this fair?

  4. Reproducibility is also a concern. If the authors retrain the baselines for this task, the authors should provide detailed parameter settings for comparison fairness.

  5. The evaluation does not adequately distinguish learning the diurnal cycle from predicting cloud-driven variability. Given the strong influence of solar geometry, comparisons with generic neural networks are insufficient. The authors should include previous-day persistence and clear-sky-based baselines, and report errors for daytime periods, cloud-induced attenuation, localized regions, and rapid fluctuations.

Questions for authors
  1. Please clarify which days and months the data cover.
  2. What is the difference with AlphaPre, and also include this as a new baseline.
  3. After accounting for solar geometry and the diurnal cycle, how much additional gain remains in predicting weather anomalies?
  4. How do the authors apply LSTM for the spatiotemporal prediction task?
Reviewer_sUU9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose SolarSpecNet, a method to forecast 20 hourly frames of ERA5 surface solar radiation fields from 5 input hours. It has three parts: Structural Evolution Branch (SEB): predicts future Fourier phase from past phase maps plus frequency-coordinate maps, then pairs it with the last observed amplitude. Radiative Energy Branch (REB): an FNO-like spatial/spectral network that predicts the fields directly. Fusion module: concatenates the SEB output, the REB output, and a low-frequency amplitude–phase recombination, then refines them with a small CNN. On a test set of 120 windows, it achieves the lowest MSE, MAE, and RMSE and the highest PSNR among six baselines (statistical, recurrent, convolutional, spectral, and foundation models). At the same time, ICNO is better on SSIM and LPIPS. Ablations remove each module and each auxiliary loss.

Critical strengths

The paper applies phase-guided structure modeling (originally from AlphaPre, CVPR 2025) to gridded ERA5 surface radiation. This is a plausible fit, since the moving day–night terminator is largely a translation captured by Fourier phase. In addition, forecasting full gridded radiation fields, rather than station series, matches the needs of distributed PV and regional grid planning. Beyond pixel metrics, the paper examines radial power spectra (over-smoothing), regional-mean trajectories across sunrise and sunset ramps, and error localization at the day–night terminator. This evaluation lens fits the field, though it remains largely qualitative and based on a single test sample.

Critical weaknesses

The proposed architecture closely resembles AlphaPre, an amplitude–phase disentanglement model for precipitation nowcasting, which the paper does not cite. Its phase branch, amplitude branch, fusion module, auxiliary losses, and input–output configuration correspond closely to AlphaPre’s public implementation. The target is dominated by the deterministic diurnal cycle, yet no simple reference (persistence or a diurnal/clear-sky baseline) is reported. This is compounded by a short test period drawn from a single season, results from a single training run without variance estimates, and metrics reported only in normalized space. The claimed superiority over the baselines is therefore not convincingly supported.

Questions for authors
  1. How does SolarSpecNet differ from AlphaPre (Lin et al.)? Its branches, fusion module, losses and input–output setup appear to correspond closely to AlphaPre’s public implementation. Why is AlphaPre neither cited nor used as a baseline?
  2. How does the model compare with a persistence or diurnal (clear-sky) reference?. Can you report mean ± standard deviation over several seeds?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.