Horizon-specific Expert Fusion for Photovoltaic Power Forecasting

1. Introduction

• Forecast horizons change how informative recent power is, while clouds and operating conditions create departures from the regular solar-generation envelope.The introduction frames this as a problem of processes acting on different time scales.Source: §1 · p. 2
• Recent observations and forecast weather are not uniformly reliable: NWP resolution and averaging limit its ability to resolve local fluctuations.The system must balance recent measurements against weather forecasts without extending transient disturbances too far ahead.Source: §1 · p. 2
• The study combines neural predictors, state-scaled climatology, rolling analogs, and full-feature LightGBM with horizon-specific convex weights and separate historical-error calibration.The fusion weights are fixed across prediction origins, rather than implemented as a weather-dependent gate.Source: §1 · p. 2
• The evaluation asks whether fusion beats full-feature LightGBM and a forecasting foundation model, which experts contribute after refitting, and whether benefits persist in GEFCom2014.It uses PVDAQ station 2107 and three GEFCom2014 solar zones, retaining and separately interpreting their native temporal resolutions.Source: §1 · pp. 2–3
• The contribution is a controlled assessment of fusion’s conditional value, including complete horizon reporting, common residual calibration, expert removal, repeated hourly training, and component timing.The design distinguishes learned combination weights from the stronger claim that a complex ensemble consistently beats the best individual model.Source: §1 · p. 3

2.1. Solar representation and weather information

• Solar geometry organizes the expected generation envelope, while archived forecasts and later observations support different experiments and information-availability assumptions.The PVDAQ experiment uses archived GFS forecasts under a declared initialization-based availability assumption.Source: §2.1 · p. 3

2.2. Temporal models and forecast combination

• Temporal neural models, feature-rich LightGBM, and pretrained forecasters offer distinct alternatives; horizon-specific nonnegative weights test whether their predictions contain useful complementary information.The paper applies established stacking principles rather than proposing a new stacking algorithm.Source: §2.2 · pp. 3–4

2.3. Evaluation and evidence of generalization

• PVDAQ and hourly GEFCom2014 retain their native resolutions; chronological fitting, paired date-block resampling, and repeated seeds address distinct evaluation uncertainties.A selected station’s results do not constitute a multisite evaluation.Source: §2.3 · p. 4

3.1. Forecast targets and normalization

• Targets are AC power assigned to 15-minute intervals across 16 horizons, normalized by the 893 kW DC nameplate; MAE in kW supplements the normalized score.The task scores power, not integrated energy or revenue.Source: §3.1 · p. 5
• The core history spans 96 intervals (24 hours) with observation masks; some predictors use longer causal contexts, and all measurements precede the forecast origin.The comparison matches information sources and forecast availability, although effective historical receptive fields differ.Source: §3.1 · p. 5

3.2. Astronomical representation

• Solar features use geometry calculated at interval midpoints; annual phase follows astronomical solar longitude, while civil time sets split boundaries and reporting.The representation includes clear-sky irradiance and daylight-transition information; rolling analogs retain exact 24-hour lags.Source: §3.2 · pp. 5–6

3.3. Forecast-weather representation and availability

• GFS weather inputs retain native forecast-statistic spans and availability indicators; the adapter assumes forecasts become available six hours after initialization.Future power and future observed weather are excluded; actual historical reception times are unknown.Source: §3.3 · p. 6

4. Hierarchical heterogeneous forecasting

• Each expert produces a 16-step forecast vector, with station-specific fitting, causal retrieval, horizon-specific fusion, and delayed-error calibration distinguished in the construction.The figure caption specifies archived NWP under a six-hour availability rule and calibration applied equally to evaluated models.Source: §4 · p. 6, Figure 1 · p. 7

4.1. Neural experts and residual retrieval

• Three neural adaptations provide forecasts, while a horizon-aware gate augments Cross-UNet with residuals retrieved from matching historical windows.The gate’s training-bank residuals are in-sample rather than forward out-of-fold residuals, which may affect gate generalization.Source: §4.1 · pp. 6–8

4.2. State climatology and rolling analog experts

• A train-only solar-hour climatology is state-scaled, while three rolling analog experts retrieve previously observed curves and use a climatology fallback when candidates are unavailable.Analog banks may incorporate evaluation-period history only after it has matured; they are not permanently frozen to training history.Source: §4.2 · p. 8

4.3. Full LightGBM and solar-scaled persistence features

• The separate LightGBM expert retains the full feature builder, including recent lags, rolling summaries, solar features, and weather descriptors.Its 311 columns include 170 shared, 137 horizon-specific, and four solar-derived features; the clear-sky ratio is an empirical index, not panel conversion efficiency.Source: §4.3 · pp. 8–9

4.4. Chronological two-level convex fusion

• Two successive convex combinations learn weights by horizon on chronological validation segments, then freeze them before evaluation rather than dynamically gating on weather.The optimization uses absolute-error slack variables.Source: §4.4 · p. 9

4.5. Delayed-residual calibration

• A separate state50 calibration uses matured forecast residuals, requiring at least 64 before correcting predictions and using bin-specific medians when a power bin has sufficient history.Calibration retains at most 384 residuals from the prior 30 days, with an additional 60-minute delay before a residual becomes eligible; every model starts with an empty calibrator.Source: §4.5 · pp. 9–10

5.1. Dataset, quality control, and chronological splits

• PVDAQ uses one public California station with archived GFS forecasts, chronological fitting periods, and a 2024 development replication rather than a newly untouched test set.The selected meter is on a 15-minute grid; the strongest standalone supervised comparator selected on validation data is LightGBM.Source: §5.1 · pp. 10–11

5.2. Training budgets and comparator inputs

• The study reports training budgets and inputs for neural models, full-feature LightGBM, and zero-shot and fine-tuned Chronos-2 comparators.The disclosed neural and foundation-model budgets are not FLOP-matched, and pretraining overlap with the public station is not independently known.Source: §5.2 · p. 11

5.3. Metrics, uncertainty, and decision rule

• The primary daylight metric averages normalized MAE equally across horizons, and paired block-resampling intervals are conditional on the fixed station, evaluation period, and fitted models.The intervals do not incorporate retraining uncertainty, public-period exposure, or uncertainty across sites; the predeclared decision rule also requires a 0.2-point all-horizon score gain and no 240-minute degradation.Source: §5.3 · p. 12

5.4. Expert removal and hourly adaptation

• The hourly adaptation evaluates three GEFCom2014 zones at known sites, using an input- and resolution-adapted ensemble with three training seeds and four forecast horizons.Its all-times normalized MAE is reported separately from PVDAQ daylight error; block intervals describe date variation, while seed standard deviations describe training variability.Source: §5.4 · p. 13

6.1. Aggregate and horizon-specific performance

• 4.315243% daylight nMAE compares with 4.500400% for full LightGBM and 4.592130% for fine-tuned Chronos-2; at 240 minutes, the respective values are 5.242859%, 5.422484%, and 5.663343%.Relative MAE reductions are 4.11% against LightGBM and 6.03% against Chronos-2.Source: Table 2 · p. 14, §6.1 · p. 14

6.2. Paired uncertainty and practical magnitude

• The LightGBM all-horizon gain is 0.1852 pp with interval [0.1181, 0.2470] pp, positive but below the predeclared 0.2 pp practical threshold.Paired intervals support positive conditional aggregate differences against both declared controls.Source: Table 3 · p. 15, §6.2 · p. 16

6.3. Learned fusion and effective expert weights

• Equal weighting yields daylight nMAE 5.233791%; validation weights give LightGBM 62.0% at 15 minutes and 73.9% at 240 minutes, with some experts receiving zero or very small weights.The fitted coefficients are horizon-specific, not causal feature attributions.Source: §6.3 · p. 16, Figure 3 · p. 17

6.4. Calibration and monthly stability

• Calibration changes scores by +0.0332 pp for the ensemble, +0.0401 pp for LightGBM, and −0.0005 pp for Chronos-2; monthly, fusion beats equal weighting in all ten months.The ensemble exceeds LightGBM in eight months and fine-tuned Chronos-2 in six.Source: Table 4 · p. 17, §6.4 · p. 18

6.5. Contribution and redundancy of individual experts

• Removing N-HiTS, TimeMixer, or LightGBM raises aggregate error, but only N-HiTS has a strictly positive contribution after adjustment for all eight comparisons.The state climatology and second analog expert can be removed without changing the reported aggregate result; eight predictors are not established as an optimum.Source: §6.5 · p. 18

6.6. Results on GEFCom2014 Solar

• Figure 4 retains negative monthly score gains; its aggregate comparison does not establish dominance in each season.Source: Figure 4 · p. 19
• 2.140543% versus 2.142789% mean normalized MAE for the ensemble and LightGBM, respectively, with no clear evidence of improvement over LightGBM.The paired interval for their difference is [−0.035991, 0.030584].Source: §6.6 · p. 18, Table 6 · p. 20
• At four hours, learned fusion improves on equal weighting by 0.334148 percentage points, with interval [0.151189, 0.521688].Across the hourly experiment, its improvement over equal weighting is 0.268626 percentage points, with interval [0.093798, 0.454062].Source: §6.6 · pp. 18, 20

6.7. Computational requirements and numerical consistency

• On an NVIDIA RTX A6000, the 16 LightGBM predictors take 0.338 ms per origin after feature construction, but these component timings are not an end-to-end service benchmark.The measurements exclude model loading, weather acquisition, raw-file parsing, and residual-history maintenance.Source: §6.7 · pp. 20–21

6.8. Interpretation and practical implications

• Across both public sources, learned fusion improves on equal weighting, while its improvement over LightGBM is positive on PVDAQ but negligible on hourly GEFCom.The authors note that additional experts are useful only insofar as validation identifies complementary information.Source: §6.8 · p. 21
• Expert removal indicates that some historical components can be removed without measurable loss, while the adjusted N-HiTS interval indicates a contribution from at least one temporal predictor.Source: §6.8 · p. 21

6.9. Limitations

• The inference is limited by one PVDAQ station, GEFCom’s June-only hourly evaluation, explicit weather-availability assumptions, and training-bank rather than forward out-of-fold residuals.Source: §6.9 · p. 22

7. Conclusions

• On PVDAQ station 2107, the ensemble attains daylight normalized MAE of 4.315%, reducing error by 4.11% against full-feature LightGBM and 6.03% against fine-tuned Chronos-2 under common calibration.The gain over LightGBM is positive under paired resampling but remains below the predefined practical threshold.Source: §7 · p. 22
• Across three training seeds, the GEFCom2014 adaptation improves on equal weighting while performing comparably to LightGBM.Source: §7 · p. 22
• Expert removal shows that several components are redundant; the results support horizon-specific combinations while indicating that a larger ensemble’s value depends on measurable complementarity.The authors identify reduced configurations and prospective multisite evaluation as next steps toward assessing operational usefulness.Source: §7 · p. 23

Data availability

• PVDAQ measurements and metadata are available from the Open Energy Data Initiative, historical GFS forecasts from the NCAR Geoscience Data Exchange, and GEFCom2014 Solar is described by Hong et al..Derived metric tables, removal contrasts, configurations, and timing results accompany the source package; a public implementation repository remains to be established.Source: Data availability · p. 23

Appendix A. Complete horizon results

• The appendix reports daylight and all-times nMAE by horizon, and notes that nighttime targets make all-times error smaller without constituting a separate model improvement.Source: Appendix A · pp. 23–24

• Appendix B.1 reports ensemble-minus-comparator score differences, in percentage points, for every evaluation month.Source: Appendix B · p. 24
• The month table is descriptive and uses the same common calibration.Source: Appendix B · p. 24
• The listed comparators are LightGBM, Chronos-2 FT, and equal weights.Source: Appendix B · p. 24