PaperlayerView exact revision on arXiv

Source-linked AI summary

GateFormer: Forecast-Space Decomposition for Intra-Hour Solar Irradiance Prediction from Sky Images

David Hamlyn, Sunny Chaudhary, Jonathon Hare, Tasmiat Rahman

arXiv:2610.30942v1eess.IVcs.AIcs.CVcs.LG

TL;DR

Sub-hourly irradiance forecasting must anticipate rapid solar variability, and sky-image forecasts need to combine cloud dynamics with recent irradiance state. GateFormer uses a shared GHI-conditioned visual representation to produce complementary forecasts and blend them in prediction space. It achieves the lowest final-horizon RMSE among evaluated baselines across three sites, with forecast skill of 22.97–28.86% relative to smart persistence.

  • Problem

    Intra-hour forecasting must combine visual cloud information with current irradiance state, while representing rapid changes that challenge a single forecast pathway.

    Source: §1 · p. 1

  • Method

    GateFormer embeds observed GHI into visual tokens, predicts a trajectory and final-horizon change from shared features, then adaptively blends the forecasts and predicts a supervised residual correction.

    Source: §1 · p. 2, §3.1 · p. 3, §3.2 · p. 4

  • Results

    Across three sites, GateFormer achieves the lowest final-horizon RMSE among evaluated baselines, with forecast skill of 22.97–28.86% relative to smart persistence.

    Source: §6 · p. 8, §7 · p. 12

  • Takeaways & Limitations

    The results support forecast-space decomposition for short-term PV balancing, storage scheduling, and ramp-rate management using sky images and measured GHI.

    Source: §6.3 · p. 11, §7 · p. 12

    Residual correction improves positive ramps but can degrade negative-ramp cases, where its direction accuracy is weaker.

    Source: §6.3 · p. 11

Abstract

from arXiv · show

Sub-hourly dispatch, balancing, and storage control increasingly depend on forecasts that anticipate rapid variability in solar generation. Accurate sky-image-based irradiance forecasting requires a combination of visual cloud dynamics with recent irradiance state while adapting to different short-term irradiance-change regimes. To address this, we introduce GateFormer, a GHI-conditioned dual-decoder ViT for intra-hour global horizontal irradiance (GHI) forecasting through forecast-space decomposition. GateFormer embeds observed GHI into the visual token stream, allowing interaction with sky-image patch features through self-attention. From this shared representation, a Trajectory Head predicts future GHI sequence, while a Delta Head predicts final-horizon change from the latest observed GHI. The two final-horizon forecasts are adaptively weighted in prediction space, with magnitude- and direction-supervision targeting the remaining error. On the NREL-TSI benchmark, GateFormer attains 125.01 W/m^2 RMSE at +30 min, improving over a state-of-the-art transformer baseline by 9.6% under paired-timestamp evaluation. Across NREL-TSI, Chilbolton, and SIRTA, GateFormer achieves the lowest final-horizon RMSE among the evaluated baselines, with forecast skill of 22.97-28.86% relative to smart persistence. Component ablations show adaptive weighting outperforms both individual decoders and fixed averaging, distinguishing GateFormer from conventional forecast combination, while the supervised correction further reduces RMSE by 1.68 W/m^2 and MAE by 4.86 W/m^2. Ramp-regime diagnostics show larger gains during operationally important positive ramps, reducing RMSE by 26.27 W/m^2. These results identify forecast-space decomposition as an alternative forecasting formulation, with potential relevance to other multimodal short-horizon forecasting tasks and deployment using only sky images and measured GHI.

1. Introduction

Sub-hourly irradiance forecasts matter for grid operations, but rapid cloud-driven changes challenge forecasting at these horizons. GateFormer addresses this by combining observed GHI with sky-image features and adaptively blending complementary forecasts.

1. Introduction

  • Sub-hourly GHI forecasts support grid balancing, storage scheduling, ramp-rate control, and short-term energy management as power systems move toward finer-grained operation.NWP often falls short at this resolution because typical operational model outputs do not capture the relevant spatial and temporal scales.Source: §1 · p. 1
  • Accurate intra-hour forecasting requires interpreting visual cloud information together with recent irradiance state, especially when cloud motion drives rapid departures from persistence-like behavior.The motivation is a forecasting formulation with distinct predictions of future irradiance state and a learned combination of those predictions.Source: §1 · p. 1
  • GateFormer embeds observed GHI into visual tokens, then uses Trajectory and Delta decoders to predict the future sequence and final-horizon displacement from a shared representation.A learned prediction-space blend combines the two final-horizon forecasts, while a residual head predicts the remaining error using magnitude and direction supervision.Source: §1 · p. 2
  • Unlike conventional ensembles or MoE routing, GateFormer forms both candidate forecasts from one shared representation and gates them only after explicit irradiance predictions are produced.The design targets both gradual trajectory evolution and direct final-horizon displacement.Source: §1 · p. 2

2. Related Work

Prior work uses sky images, multimodal fusion, specialized pathways, and forecast combination for solar prediction. GateFormer’s distinction is combining alternative irradiance forecasts generated from a shared visual–irradiance representation.

2. Related Work

  • Classical sky-image forecasting uses cloud segmentation, optical flow, and tracking, but depends on preprocessing and assumptions about cloud motion, height, and transmittance.Deep learning instead learns predictive visual representations directly from imagery.Source: §2 · p. 2
  • Transformer and multimodal methods incorporate sky imagery, irradiance history, or clear-sky information, but still optimize a single forecast mapping rather than complementary objectives.Prior work reports improved irradiance or ramp-event performance, including Liu et al.’s 4.28% nRMSE at 10 minutes on NREL SRRL and Zhang et al.’s ramp-event balanced-accuracy gains.Source: §2 · p. 2
  • Other specialized architectures separate modalities, processing stages, or deterministic and probabilistic dynamics, rather than generating and combining alternative irradiance forecasts from one latent representation.Retaining both candidate predictions exposes their values and disagreement to the gate before the final prediction is formed.Source: §2 · p. 3
  • Forecast-level combination and input-dependent weighting have prior precedents, including solar forecasts blended from separate ASI, satellite, and persistence systems.GateFormer addresses the unresolved question of whether a shared visual–irradiance representation can support distinct forecast objectives, conditional final-horizon weighting, and supervised correction of post-blend error.Source: §2 · p. 3

3. Method

GateFormer encodes sky images and observed GHI in a shared representation, produces complementary forecasts, and adaptively blends and corrects their final-horizon predictions.

3.1. Overall Architecture

  • The shared encoder inserts observed GHI as a learned token in each image’s patch-token stream, then pools frame representations and temporal differences into memory for both forecasting heads.Inputs include four past sky images, corresponding GHI anchor values, and temporal/clear-sky features.Source: §3.1 · pp. 3–4

3.2. Dual Heads

  • The Trajectory Head predicts the full future GHI sequence, while the Delta Head predicts signed final-horizon change from the latest observed GHI.The Delta Head adds its predicted change to the latest observation; the adaptive blend determines each sample’s relative reliance on the two forecasts.Source: §3.2 · p. 4

3.3. Soft Gating Mechanism

  • The adaptive blend combines the two final-horizon forecasts into a Base forecast, and a residual head predicts the remaining post-blend error.The blend weight is learned through the final prediction objective and residual-supervision terms, not directly supervised to select the locally better decoder.Source: §3.3 · p. 5

3.4. Loss Functions

  • Training proceeds in two stages: decoder pretraining first, then freezing the encoder and forecasting heads while training the blend-residual module.The gate objective includes final-prediction, residual-error, and residual-direction losses; event loss is zero when a minibatch contains no qualifying event.Source: §3.4 · pp. 5–6

4. Dataset

GateFormer is evaluated on three sky-imaging datasets with co-located GHI measurements, using each site's native cadence and a common final forecast step. The data are split by days within each month, with paired image-radiometry timestamps and daylight filtering.

4. Dataset

  • The evaluation uses NREL-TSI at 10-minute cadence, Chilbolton at 5-minute cadence, and SIRTA at 2-minute cadence.The datasets contain 68,740, 158,461, and 365,811 sequences, respectively.Source: §4 · p. 6
  • Sequences use four past frames and three future horizons, with the final forecast step at +30 min for NREL-TSI, +15 min for Chilbolton, and +6 min for SIRTA.The intermediate horizons provide trajectory supervision while the final step is shared by the Delta Head and adaptive blend.Source: §4 · p. 6
  • The split assigns days 5–9 of each month to testing, days 15–19 to validation, and remaining eligible days to training.Splits are generated before sequence construction, with no shared images or GHI target timestamps across subsets.Source: §4 · p. 6
  • Image-radiometry pairing uses a ±10s tolerance, drops sequences with missing frames, and retains daylight targets above 10 W/m² while excluding days with maximum GHI below 50 W/m².Source: §4 · p. 6
  • Images are resized to 224 by 224 pixels and ImageNet-standardised, without rotation, colour perturbation, blurring, or similar augmentation.The authors identify richer sky-specific augmentation as future work.Source: §4 · p. 6

5. Implementation and Evaluation Metrics

The study reports MAE, RMSE, and forecast skill under paired-timestamp evaluation against clear-sky-scaled persistence. GateFormer uses a pretrained ViT with a two-stage training procedure, and its architecture and settings are held fixed across the three sites.

5. Implementation and Evaluation Metrics

  • Evaluation reports MAE, RMSE, and FS against clear-sky-scaled persistence using paired-timestamp evaluation.Source: §5 · p. 6
  • Simplified Solis clear-sky irradiance is used only to construct the shared persistence reference and does not affect GateFormer predictions.Source: §5 · p. 7
  • The DeiT-III Medium ViT backbone uses 16 × 16 image patches, two forecast decoders, and a compact gating MLP.The backbone is pretrained on ImageNet.Source: §5 · p. 7
  • Training first pretrains the shared encoder and both decoders for 4 epochs, then freezes them while training the blend-residual module for 6 epochs.Source: §5 · p. 7
  • The Delta loss weights change magnitude and adds an event-asymmetric penalty for large upward under-prediction and downward over-prediction.The event threshold is 150 W/m²; the Delta loss uses change-magnitude weighting of 2, scale 200 W/m², and cap 3.Source: §5 · p. 7
  • Hyperparameters are selected using NREL-TSI validation RMSE and then fixed across NREL-TSI, Chilbolton, and SIRTA without site-specific tuning.Source: §5 · p. 7

6. Results and Discussion

GateFormer has the lowest final-horizon RMSE among evaluated baselines across three sites, with the largest reported component gains coming from adaptive blending. Seasonal and ramp-regime analyses show where its residual correction helps and where it can degrade performance.

6. Results and Discussion

  • At NREL-TSI, GateFormer reaches 125.01 W/m² Step 3 RMSE versus 138.35 W/m² for SIAT; across three sites, it has the lowest final-horizon RMSE and 22.97%–28.86% forecast skill.Final-horizon SIAT RMSE reductions reach 9.64%.Source: Table 3 · p. 8, §6 · p. 8

6.1. Seasonal Performance

  • Across seasons, the final forecast has the best seasonal RMSE, residual gains remain positive, and Base and final forecast skill varies by only 2–3%.In JJA, residual correction improves RMSE by −2.22 W/m² and MAE by −7.64 W/m²; the smallest RMSE gain is −0.83 W/m² in MAM.Source: Table 4 · p. 8, §6.1 · p. 8

6.2. Ablations and Analysis

  • The adaptive blend reaches 126.69 W/m² RMSE, beating equal averaging by 2.93 W/m², while the Delta Head is strongest in extreme-change and large downward-ramp regimes.Equal averaging reaches 129.62 W/m², versus 129.23 W/m² for the Trajectory Head; for drops exceeding 250 W/m², Delta Head RMSE is 210.10 W/m² versus 296.10 W/m² for Trajectory Head.Source: §6.2 · p. 9, Table 6 · p. 9
  • Replacing the learned Delta Head with persistence leaves only a small additional residual benefit, with RMSE changing from 128.23 W/m² to 128.00 W/m².The authors interpret this as evidence that the learned branch supplies complementary image-informed structure for blending and correction.Source: §6.2 · p. 9
  • Removing the GHI token raises push-direction accuracy from 61.39% to 68.68% but reduces positive-ramp residual gains and worsens negative-ramp degradation.For rises above 250 W/m², residual gain changes from 26.27 to 24.19 W/m²; for drops below 250 W/m², degradation changes from 15.46 to 26.33 W/m².Source: §6.2 · p. 9
  • The paper configuration has the best tested overall RMSE, but loss ablations trade aggregate and positive-ramp performance against negative-ramp RMSE; predicting Delta at additional horizons increases final-horizon error.The tested configuration was retained across datasets, with the Delta Head serving as a final-horizon specialist while the Trajectory Head supplies intermediate forecasts.Source: §6.2 · p. 10, Table 7 · p. 10, Table 8 · p. 10
  • GateFormer Base and Final outperform parameter-matched single-head controls and independently trained post-hoc combinations; Final reduces RMSE by 1.93 W/m² versus the strongest conventional combination.Relative to the stronger final-only single-head control, Base reduces RMSE by 3.13 W/m² and Final by 4.81 W/m².Source: Table 10 · p. 10, §6.2 · pp. 10–11

6.3. Ramp-Regime Diagnostics

  • Residual correction improves positive-ramp RMSE by 26.27 W/m² for rises above 250 W/m² but can degrade negative ramps; deployment timing and lower-resource hardware were not evaluated.The framework adds dual-decoder and second-stage training cost, while inference uses one shared-backbone pass followed by lightweight gating and residual operations.Source: Table 11 · p. 11, §6.3 · p. 11

7. Conclusion

GateFormer combines GHI-conditioned visual encoding, complementary forecast decoders, adaptive blending, and residual correction. It achieves the lowest final-horizon RMSE among evaluated baselines across three sites, with gains during positive ramps.

7. Conclusion

  • GateFormer embeds observed GHI into the transformer token stream, adaptively blends two complementary forecast decoders, then applies residual correction.The shared representation supports the two forecast decoders; the conclusion identifies visual-irradiance encoding, trajectory prediction, final-horizon change, adaptive blending, and residual correction as the stages.Source: §7 · pp. 11–12
  • Across NREL-TSI, Chilbolton, and SIRTA, GateFormer achieves the lowest final-horizon RMSE among evaluated baselines, with 22.97%–28.86% FS; at NREL-TSI +30 min, it reaches 125.01 W/m² RMSE, 9.6% better than SIAT.The NREL-TSI comparison uses paired-timestamp evaluation.Source: §7 · p. 12
  • Residual correction reduces RMSE by up to 26.27 W/m² during positive ramps, while negative-ramp correction remains an area for improvement.The paper identifies short-term PV balancing, storage scheduling, and ramp-rate management under rapidly changing cloud conditions as relevant applications.Source: §7 · p. 12

CRediT Authorship Contribution Statement

The statement assigns writing, software, supervision, conceptualization, visualization, and investigation roles across four authors.

CRediT Authorship Contribution Statement

Related papers

Loading 2610.30942v1…