Source-linked AI summary
IMPACT: Inverter-Informed Multimodal Learning for Photovoltaic Generation Forecasting
Anushka Katiyar, Cong Feng, Jing Wang, Tristan Liu, Simon Julien, Zachary Jacobs
TL;DR
Intra-hour PV forecasting is challenging because cloud dynamics can shift output rapidly, while existing approaches have gaps in inverter-scale spatial information and multimodal integration. IMPACT combines four input branches in a staged architecture and corrects a smart-persistence forecast. It outperformed all 14 baselines, leading the strongest by 0.099 FSS, and each branch improved skill in ablation.
Problem
Fast cloud dynamics complicate intra-hour PV forecasting, and prior approaches often lack inverter-level spatial context or staged multimodal integration.
Source: §I · p. 1Method
IMPACT combines temporal history, inverter telemetry, horizon covariates, and meteorological snapshots through gated cross-attention, then predicts corrections to smart persistence.
Source: §I · p. 1, B. Branch Encoders · p. 2, C. Fusion · p. 3Results
IMPACT outperformed all 14 baselines, leading XGBoost-Q by 0.099 FSS; every added branch improved forecast skill.
Source: B. Probabilistic Forecasting Performance · p. 4, §V · p. 5, C. Sensitivity Analysis · p. 5Takeaways & Limitations
Probabilistic intra-hour PV forecasts support dispatch, reserve procurement, and market decisions when cloud-driven output variability must be managed.
Source: §I · p. 1The current fully data-driven inverter adjacency lacks a physically grounded prior based on true inverter geographic coordinates.
Source: §V · p. 5
Abstract
from arXiv · showhide
This work develops a physics-informed multimodal deep learning framework for 10- to 110-minute-ahead probabilistic photovoltaic power forecasting. The forecasting model is built on a four-branch architecture that separately encodes plant-level temporal history and inverter-level spatial cloud dynamics through a graph attention encoder with a learned adjacency, known-future solar geometry, and the current meteorological state, fused through gated cross-attention. The geometry branch adds two forward-looking physical signals, spatial cloud heterogeneity and inverter ramp-down activity, each empirically weighted toward the horizon where it is most predictive. The decoder anchors its prediction to a smart persistence baseline, learning a horizon-conditioned correction, and the resulting quantile forecasts are calibrated online through a kernel-weighted adaptive conformal calibration step. The proposed model outperforms all 14 baselines, achieving a forecast skill score (FSS) of 0.2950, 0.099 above the next-best model. A branch-composition ablation shows that every branch contributes positively and the model retains substantial skill in challenging overcast and variable sky conditions (FSS 0.198 and 0.255).
I. INTRODUCTION
Intra-hour PV forecasts matter for grid and market decisions, but fast cloud dynamics make them difficult. IMPACT addresses gaps in spatial inverter information and multimodal fusion with a four-input forecasting model.
I. INTRODUCTION
- Fast cloud motion and localized atmospheric dynamics can sharply shift PV output within minutes, complicating intra-hour forecasting.Data-driven models may lack physical grounding, while physics-based models struggle with fast-moving transient clouds.Source: §I · p. 1
- Probabilistic forecasts support reserve, bidding, and real-time operating decisions by expressing likely outcomes rather than only expected output.They can indicate whether flexibility is likely to persist across a commitment period.Source: §I · p. 1
- Prior approaches often lack inverter-level spatial cloud context, staged fusion across modalities, or integration of telemetry into a broader intra-hour probabilistic framework.The cited work describes these as three gaps addressed by IMPACT.Source: §I · p. 1
- IMPACT combines inverter telemetry, temporal history, solar geometry, and meteorological snapshots through gated cross-attention, with predictions anchored to smart persistence.Source: §I · p. 1
II. METHODOLOGY
The methodology predicts quantiles across 11 intra-hour horizons using four encoders and a smart-persistence anchor. Its inputs include plant and inverter histories, horizon covariates, and meteorological measurements.
II. METHODOLOGY
- IMPACT predicts site-level AC power quantiles at 11 horizons from 10 to 110 minutes, correcting a smart-persistence benchmark before online calibration.Source: §II · p. 1
A. Data
- Experiments use 10-minute data from a 205-MW plant with 112 inverters, four meteorological stations, and three expanding-window validation splits.The data span January 2020 through October 2023; sky regimes are classified as clear, variable, or overcast.Source: A. Data · p. 2
B. Branch Encoders
- Four encoders represent temporal history, inverter-block dynamics, horizon covariates, and meteorological snapshots in a shared 96-dimensional space.The inverter branch uses convolutional and recurrent processing followed by graph attention with learned adjacency; the horizon branch weights spatial variability and ramp-down signals toward their predictive lead times.Source: B. Branch Encoders · pp. 2–3
C. Fusion
Fusion is staged according to the structure of the inputs: temporal and inverter sequences interact, meteorology modulates their combined representation, and horizon covariates bypass fusion.
C. Fusion
- Temporal and inverter sequences combine through cross-attention, while meteorological features gate their fused representation and horizon covariates are deferred to the decoder.Source: C. Fusion · p. 3
D. Decoding and Probabilistic Output
The decoder uses horizon-specific queries to produce quantile corrections to a smart-persistence forecast, then constrains and orders the resulting outputs.
D. Decoding and Probabilistic Output
- The decoder predicts 99 quantile corrections to smart persistence, scales them with a horizon-conditioned gate, clips predictions to the valid power range, and sorts outputs by quantile.Horizon queries cross-attend to the fused sequence, receive a meteorological-branch bias, and are mapped to corrections by an MLP.Source: D. Decoding and Probabilistic Output · p. 3
E. Loss and Online Calibration
The model is trained with a blend of pinball loss and CRPS, then calibrated online by adjusting forecasts using similar past residuals. Calibration is performed separately for each horizon and quantile to address under-coverage.
E. Loss and Online Calibration
- Online calibration adjusts under-covering raw outputs separately for each horizon and quantile using residuals stored in a rolling buffer.The buffer includes a feature vector describing current conditions, including time of day, solar zenith, and recent CSI variability.Source: E. Loss and Online Calibration · p. 3
- Training uses an equal blend of pinball loss and continuous ranked probability score (CRPS).Source: E. Loss and Online Calibration · p. 3
- Each buffered residual is weighted by its similarity to current conditions, and the correction is the weighted residual quantile at the nominal quantile level.The weights are normalized to sum to 1; the procedure builds on sequential predictive conformal inference.Source: E. Loss and Online Calibration · p. 3
III. EXPERIMENTAL SETUP
Evaluation compares median-forecast accuracy and probabilistic forecast quality against the smart-persistence benchmark, alongside 14 baselines spanning major forecasting paradigms. Probabilistic metrics assess the calibrated predictive distribution, while interval metrics are evaluated at four representative coverage levels.
A. Evaluation Metrics
- Forecast skill relative to SP-PI is computed from RMSE for median forecasts and from CRPS for calibrated probabilistic forecasts.Coverage error and interval width are aggregated across the 80%, 90%, 95%, and 98% levels.Source: A. Evaluation Metrics · p. 3
B. Baselines
- The 14-baseline suite spans smart persistence, statistical and classical methods, linear models, deep learning, and transformers, evaluated on identical splits, preprocessing, and loss functions.Smart persistence serves as the domain-standard floor; the suite also tests whether attention’s added complexity is required.Source: B. Baselines · p. 3
C. Ablation Study
The ablation isolates the contribution of the spatial-inverter and horizon-covariate branches through a cumulative branch-composition comparison. Each added branch retains its full mechanism, and the comparison is run across all cross-validation folds.
C. Ablation Study
- The branch-composition ablation adds branches cumulatively from B1 through the full model across all cross-validation folds.Each added branch retains its full mechanism.Source: C. Ablation Study · pp. 3–4
IV. RESULTS
IMPACT achieves the strongest deterministic accuracy among the evaluated models, outperforming SP-PI and all 13 other baselines on the reported aggregate metrics. Its overall forecast skill is Sf = 0.1616.
A. Deterministic Forecasting Performance
- 0.1375 nRMSE and 0.0669 nMAE are IMPACT’s lowest values, while its R2 of 0.7683 is highest among evaluated models, leading XGBoost-Q by 0.008 nRMSE and 0.024 R2.Against SP-PI, IMPACT reduces nRMSE from 0.1640 to 0.1375 and nMAE from 0.0858 to 0.0669, and increases R2 from 0.6674 to 0.7683; overall forecast skill is Sf = 0.1616.Source: A. Deterministic Forecasting Performance · p. 4
B. Probabilistic Forecasting Performance
IMPACT leads the probabilistic forecasting comparisons across horizons, while its performance varies across sky conditions, seasons, and times of day. Ablations find positive skill gains from every added branch, with the largest gain from Branch 3’s horizon-targeted spatial feature.
B. Probabilistic Forecasting Performance
- IMPACT leads the strongest baseline, XGBoost-Q, by 0.099 FSS and remains top-ranked with positive skill across all 10- to 110-minute horizons.At 10 minutes, every learned baseline scores below zero; most baselines do not turn positive until 30 minutes.Source: B. Probabilistic Forecasting Performance · p. 4
- Fig. 2 compares actual power, IMPACT, prediction intervals, and four baselines at 10, 60, and 110 minutes across three sky conditions.The reported gap is widest at 10 minutes and narrows at 60 and 110 minutes as baselines catch up.Source: Figure 2 · p. 4, B. Probabilistic Forecasting Performance · p. 4
- Fig. 3 presents per-quantile coverage against the perfect-calibration diagonal and mean PINAW at nominal 80%, 90%, 95%, and 98% levels.The paper reports strong reliability and sharpness across all horizons.Source: B. Probabilistic Forecasting Performance · p. 4, Figure 3 · p. 5
C. Sensitivity Analysis
- Every added branch improves forecast skill, and Branch 3’s horizon-targeted spatial feature contributes the largest single gain, +0.038 FSS.The ablation uses a nine-quantile configuration to construct the four operational prediction intervals.Source: C. Sensitivity Analysis · p. 5, Table II · p. 5
- Skill varies by sky condition, season, and time of day: clear skies perform best, overcast worst, summer leads, and transition windows outperform midday.Overcast and variable sky have MACE values of 0.023 and 0.022; midday is best calibrated by MACE at 0.007, while evening is sharpest and morning has the widest intervals.Source: C. Sensitivity Analysis · p. 5, Table III · p. 5
V. CONCLUSION AND FUTURE WORK
IMPACT outperforms all 14 baselines, with the strongest baseline trailing by 0.099 FSS. The authors identify external irradiance products, inverter coordinates, and further investigation of performance gaps and transition-window asymmetry as future work.
V. CONCLUSION AND FUTURE WORK
- IMPACT outperforms all 14 baselines, leading the strongest by 0.099 FSS; every branch contributes positive, nonredundant skill, with spatial and horizon-covariate branches contributing most.Overcast conditions are hardest, with variable conditions close behind; calibration asymmetry between morning and evening transitions remains unexplained.Source: §V · p. 5
- Future work includes combining IMPACT forecasts with external vendor irradiance products to extend accuracy across the full horizon range.The authors also identify performance gaps and transition-window asymmetry for further investigation.Source: §V · p. 5
- True inverter geographic coordinates could provide a physically grounded adjacency prior, which the current fully data-driven adjacency lacks.Source: §V · p. 5