PaperlayerView exact revision on arXiv

Source-linked AI summary

Constructed regional photovoltaic generation fields improve ultra-short-term distributed solar forecasting

Dayin Chen, Xiaodan Shi, Xin Xu, Zhongyue Su, Shibo Zhu, Dongxiao Zhang, Yuntian Chen, Jinyue Yan

arXiv:2610.16110v1cs.AIcs.CE

TL;DR

Dense PV fleets provide measurements of regional generation dynamics, but forecasting remains difficult amid local weather variability and heterogeneous system responses. SatPV-Field predicts spatially structured regional PV generation fields and uses them to guide station forecasts. It reduces station-level MAE against both a single-station temporal baseline and a matched graph-based approach.

  • Problem

    Distributed PV measurements reveal regional generation dynamics, but explicit regional generation states are not fully characterized by meteorological information alone.

    Source: §I · pp. 1–2, G. Discussion and Practical Implications · p. 11

  • Method

    SatPV-Field predicts regional PV generation fields from historical fields, satellite-derived future shortwave radiation guidance, and meteorological forecasts, then uses predicted fields to guide station predictions.

    Source: §IV · p. 11

  • Results

    Approximately 10.0% lower station-level MAE than the best-performing single-station temporal baseline and 2.5% lower than its graph-based counterpart.

    Source: §IV · p. 11

  • Takeaways & Limitations

    The findings establish a regional-to-local forecasting pathway connecting predicted regional generation states with location-specific station predictions.

    Source: §IV · p. 11

    The evaluation covers a single regional PV fleet, and real-time deployment must account for satellite-product latency.

    Source: G. Discussion and Practical Implications · p. 11

Abstract

from arXiv · show

Ultra-short-term forecasting of densely distributed photovoltaic (PV) systems remains challenging because external meteorological and satellite products may not adequately resolve rapid local variations, while heterogeneous PV systems exhibit different generation responses under similar atmospheric conditions. Meanwhile, dense PV fleets provide spatially distributed measurements that reveal regional generation dynamics. Here, we reformulate fleet-level PV forecasting from modeling discrete station time series to predicting explicit, spatially structured regional PV generation fields. We develop SatPV-Field, a regional-to-local forecasting framework that transforms distributed PV measurements into image-like generation fields and models their spatiotemporal evolution. By combining historical PV fields with satellite-derived future shortwave radiation guidance and meteorological forecasts, the framework predicts future regional generation states, which subsequently provide location-specific guidance for individual-station forecasting. Experiments involving 496 distributed PV systems, including 446 source stations and 50 held-out target stations, demonstrate that these regional fields can be effectively forecast and used to improve station-level predictions. The proposed approach reduces station-level MAE by approximately 10.0% relative to the best-performing single-station temporal baseline and by 2.5% relative to a graph-based regional forecasting approach using the same source fleet. These findings establish an explicit spatial-field modeling paradigm that connects regional generation dynamics with individual-site forecasting.

I. INTRODUCTION

Distributed PV forecasting must account for local weather variability and heterogeneous system responses, while fleet measurements reveal regional generation dynamics. SatPV-Field forecasts explicit regional PV fields to guide station predictions, reducing MAE against temporal and graph-based baselines.

I. INTRODUCTION

  • External weather products may miss fine-scale local variability, and similar atmospheric conditions can yield different PV outputs because of site-specific system characteristics.The passage identifies limits in satellite observations and NWP forecasts, alongside variation in configuration, orientation, shading, efficiency, degradation, and operating conditions.Source: §I · p. 1
  • Distributed PV measurements reveal evolving regional generation dynamics that cannot be captured by any single site’s history.Collected fleet-wide, power measurements act as a regional sensing resource reflecting local responses to atmospheric conditions.Source: §I · pp. 1–2
  • The paper represents source-fleet measurements as spatially structured regional PV fields, making geographic relationships and temporal evolution directly modelable.This representation captures realized PV responses rather than atmospheric or irradiance conditions alone.Source: §I · pp. 2–3
  • SatPV-Field combines historical PV fields, satellite-derived future irradiance guidance, and meteorological forecasts to predict regional states that guide station-level forecasts.Target-site observations are retained for local forecasting but excluded from regional-field construction and field-level supervision.Source: §I · pp. 2–3
  • Approximately 10.0% lower station-level MAE than the best-performing single-station temporal baseline and 2.5% lower than a matched graph-based regional approach.Evaluation used 496 systems: 446 source stations and 50 held-out target stations.Source: §I · p. 3

II. METHODOLOGY

The methodology formulates fleet forecasting as prediction of spatially structured PV generation fields, then uses predicted local field context alongside target-site information for station forecasting.

A. Problem Formulation: Explicit Regional PV Generation Field Forecasting

  • Source-station power measurements are organized into a spatially structured regional PV generation field, with a support mask identifying observed grid cells.The field grid is 64×64 in this study.Source: A. Problem Formulation: Explicit… · p. 3
  • The field forecaster predicts future regional PV fields from historical fields and support masks, predicted future irradiance guidance, and future meteorological and solar-condition information.The formulation uses k historical and H future time steps, with discrete 15-minute steps.Source: A. Problem Formulation: Explicit… · p. 3
  • For each target station, location-specific predicted-field context and its support mask are combined with historical observations and future meteorological information to predict power output.The context is extracted around the target station’s location.Source: A. Problem Formulation: Explicit… · pp. 3–4

B. Source-Supported Regional PV Generation Field Construction

The framework constructs source-supported regional PV fields and combines their history with satellite-derived irradiance and meteorological information to forecast regional generation. Satellite observations provide broader atmospheric context for the PV region.

B. Source-Supported Regional PV Generation Field Construction

  • Source-station measurements are normalized, spatially aggregated into grid fields, and paired with a support mask; target stations are excluded from construction and supervision.Capacity normalization removes installed-size differences, while station-specific quantile scaling aligns fleet magnitudes without removing temporal variation.Source: B. Source-Supported Regional PV… · p. 4

C. Future Regional PV Generation Field Forecasting

  • Broader satellite observations complement limited PV-region coverage by supplying information on atmospheric evolution beyond the observed field boundaries.The resulting irradiance forecasts are combined with historical PV fields and future meteorological information to predict regional generation evolution.Source: C. Future Regional PV Generation Field… · p. 4

1) Satellite-Derived Future Irradiance Guidance:

  • Four hourly satellite observations predict three hourly shortwave-radiation fields, which are reconstructed from clear-sky-normalized ratios and cropped to guide PV-field forecasting.The predicted fields correspond to T, T + 1 h, and T + 2 h.Source: 1) Satellite-Derived Future Irradiance… · p. 4

2) Temporal Alignment of Irradiance Guidance:

Hourly irradiance guidance is aligned to the PV-field model’s 15-minute forecast resolution through fixed linear interpolation.

2) Temporal Alignment of Irradiance Guidance:

  • Fixed linear interpolation between hourly-mean guidance centers converts irradiance forecasts into an eight-step 15-minute sequence without trainable parameters.The four anchor centers are T−30, T+30, T+90, and T+150 minutes.Source: 2) Temporal Alignment of Irradiance… · pp. 4–5

3) Spatiotemporal Field Evolution Modeling:

The field forecaster combines historical PV fields with future irradiance and meteorological guidance, then predicts regional generation through complementary spatial-propagation and direct-prediction branches.

3) Spatiotemporal Field Evolution Modeling:

  • Historical PV fields, aligned irradiance guidance, and future meteorological conditions are encoded and combined to forecast regional generation evolution.The field forecaster uses a 3D convolutional encoder for historical fields and auxiliary inputs, with separate convolutional encoders for irradiance.Source: 3) Spatiotemporal Field Evolution… · p. 5
  • A displacement-based branch warps the latest field to model spatial propagation, while a direct branch captures local intensity changes that warping cannot represent.A learned pixel-wise gate adaptively combines the two predictions at each grid cell.Source: 3) Spatiotemporal Field Evolution… · p. 5
  • The field forecaster uses support-masked SmoothL1 reconstruction loss averaged over source-supported cells and forecast horizons.Source: 3) Spatiotemporal Field Evolution… · p. 5

D. Regional-to-Local Station Forecasting

Predicted regional fields provide location-specific context for target-station forecasts alongside station history and future conditions.

D. Regional-to-Local Station Forecasting

  • Each target station receives a 5×5 patch from its corresponding predicted regional field, containing generation values and a source-support mask.The latest observed support mask is reused across forecast horizons.Source: D. Regional-to-Local Station Forecasting · p. 5
  • Station forecasts combine target-station history, future meteorological and solar-condition features, and location-specific regional-field patches.A bidirectional LSTM encodes station history; separate multilayer perceptrons encode future conditions and field patches before fusion into horizon-specific forecasts.Source: D. Regional-to-Local Station Forecasting · p. 5
  • Weighted SmoothL1 training assigns weight 1.5 to horizons with a ground-truth power change of at least 0.10 p.u.; other horizons receive unit weight.Source: D. Regional-to-Local Station Forecasting · p. 5

III. EXPERIMENTS

Experiments use a Ningbo PV dataset to test regional-field and station-level forecasting with geographically separated source and target stations. The protocol combines satellite, meteorological, and PV inputs across sequential forecasting stages.

1) Dataset and Auxiliary Data:

  • The Ningbo dataset contains 496 PV systems, sampled every 15 minutes from January 2022 to June 2023; inputs include satellite products, ERA5 history, and operational weather forecasts.Weather forecasts from 12 locations are interpolated to the 64 × 64 PV-field grid; station-level prediction also uses location-specific forecasts.Source: A. Experimental Setup and Evaluation… · p. 6, 1) Dataset and Auxiliary Data · p. 6

2) Source–Target Setting and Data Split:

3) Forecasting Protocol:

  • Each forecast uses 12 historical observations to predict eight 15-minute steps, with six daily forecast origins shared by regional-field and station-level tasks.The latest PV observation is at T −15 min; satellite observations are limited to those available by issue time, without modeling product-processing or dissemination delays.Source: 3) Forecasting Protocol · p. 6
  • The irradiance and regional-field forecasters are trained first and then fixed while their predictions train the downstream station-level forecaster.The protocol yields 2,268 training, 324 validation, and 648 test forecast-origin instances.Source: 3) Forecasting Protocol · p. 6

4) Evaluation Metrics:

The evaluation measures regional-field accuracy, structure, and changes, alongside station-level accuracy. Constructed fields retain source-fleet generation characteristics across distinct operating regimes and closely match source-station observations.

4) Evaluation Metrics:

  • Regional-field evaluation uses MAE, RMSE, R2, masked global SSIM, DeltaCorr, SignAcc, and Large-∆MAE; station-level evaluation uses MAE, RMSE, and R2.Regional metrics use source-supported grid cells; station-level errors are normalized by installed capacity.Source: 4) Evaluation Metrics · p. 6
  • DeltaCorr compares predicted and observed field increments, SignAcc counts matching change signs, and Large-∆MAE measures errors where observed changes are substantial.SignAcc excludes negligible observed changes, while Large-∆MAE uses a |∆P| ≥ 0.10 threshold.Source: 4) Evaluation Metrics · p. 7

B. Validation of Regional PV Generation Field Representation

  • The three representative regimes are stable high-generation, spatially heterogeneous generation, and rapidly varying generation, shown as station measurements beside source-supported fields.The stable case has µ = 1.080, σ = 0.134, and ∆ = 0.012; the heterogeneous case has σ = 0.355; the rapidly varying case has ∆ = 0.230.Source: B. Validation of Regional PV… · p. 7, Figure 4 · p. 7
  • Across 1,907,857 station-time samples, constructed fields achieved MAE 0.0539, RMSE 0.1037, and R2 0.9025, retaining the source fleet’s main generation characteristics.Field values were sampled at source-station locations and compared with normalized observations.Source: B. Validation of Regional PV… · p. 7

C. Forecasting Performance of Regional PV Generation Fields

The proposed forecaster improves regional PV-field forecasts over persistence and U-Net baselines, with qualitative gains at longer horizons. Ablations indicate contributions from future irradiance guidance and the field-evolution design.

C. Forecasting Performance of Regional PV Generation Fields

D. Benefits of Regional Field Context for Station-Level Forecasting

Regional PV-field forecasts provide stronger station-level guidance than historical fields, irradiance guidance, or graph-based guidance; SatPV-Field also outperforms the reported temporal and graph-based benchmarks.

D. Benefits of Regional Field Context for Station-Level Forecasting

  • The comparisons use a common BiLSTM-based station backbone and target-site inputs; graph guidance uses the same 446 source stations while excluding the 50 targets.The study compares local-only, nearest-neighbor, irradiance, historical-field, predicted-field, and graph-based regional guidance.Source: D. Benefits of Regional Field Context… · p. 9
  • Predicted future PV-field guidance lowers MAE to 0.0672, improving by 7.3% over historical PV-field guidance and 5.8% over future irradiance guidance.All tested regional guidance types improve on the local-only baseline.Source: D. Benefits of Regional Field Context… · p. 9
  • Predicted PV-field guidance reduces MAE by 2.5% versus predicted graph guidance and yields lower station-wise MAE at 40 of 50 target stations.Forecasting future graph-node states reduces MAE from 0.0719 to 0.0689 versus historical graph guidance.Source: D. Benefits of Regional Field Context… · p. 9

E. Comparison with Existing Forecasting Methods

  • Among temporal baselines, BiLSTM+NWP has the lowest MAE at 0.0746, while AGCRN achieves 0.0694 among the compared graph-based models.The benchmark comparison follows a shared chronological split and forecasting protocol.Source: E. Comparison with Existing… · p. 9
  • SatPV-Field achieves the best results across all three metrics: MAE 0.0672, RMSE 0.0999, and R2 0.8266, reducing MAE by 9.9% versus BiLSTM+NWP and 3.2% versus AGCRN+NWP.The results are summarized in Table IV.Source: E. Comparison with Existing… · p. 9, Table IV · p. 9
  • Figure 7 plots station-level MAE for BiLSTM+NWP, AGCRN+NWP, and predicted future PV-field guidance, plus regional-field MAE for three field forecasters.The figure compares performance across forecast lead times.Source: Figure 7 · p. 10

F. Robustness of Regional Field Forecasting and Guidance

Predicted regional fields maintain an advantage across forecast horizons and operating conditions, while denser source fleets improve field coverage and accuracy.

1) Performance Across Forecast Horizons:

  • From 45 to 120 minutes, predicted future PV-field guidance has lower station-level MAE than the compared alternatives, with a larger advantage at medium and longer leads; its regional-field forecaster has the lowest MAE across 15–120 minutes.Persistence errors rise rapidly with lead time, while PV+NWP U-Net deteriorates more gradually.Source: 1) Performance Across Forecast Horizons · p. 10

2) Source-Fleet Density and Field Observability:

  • As source density rises from 25% to full, local-patch supported-pixel ratio increases from 0.6312 to 0.9218 and field MAE falls from 0.1251 to 0.0880.Station-level results are less sensitive: 50% and 75% density achieve similar MAEs to full density, and improvements are not strictly monotonic.Source: 2) Source-Fleet Density and Field… · p. 10

3) Performance Across Operating Conditions:

  • Predicted regional-field guidance improves station-level MAE in clear, partly cloudy, and overcast conditions, with the largest relative reduction, 8.3%, under partly cloudy conditions.It also reduces RMSE from 0.0621 to 0.0533 under low variation and from 0.1397 to 0.1304 under high variation, though errors remain larger during rapid generation changes.Source: 3) Performance Across Operating… · p. 10

G. Discussion and Practical Implications

The paper frames distributed PV measurements as an evolving regional generation state and uses predicted fields to connect fleet-level dynamics with site forecasts. It reports improved forecasting while noting limits in model design, regional scope, and satellite-product latency.

G. Discussion and Practical Implications

  • The formulation treats fleet measurements as observations of an evolving regional generation state, making spatial patterns from atmospheric conditions and heterogeneous system responses available for prediction.The framework models regional dynamics first and then uses them to inform individual-site predictions.Source: G. Discussion and Practical Implications · p. 11
  • The current regional-to-local stage uses a target-centered 5 × 5 field patch encoded by an MLP; more advanced spatial models could exploit larger contexts and multiscale interactions.The paper describes this work as an initial feasibility demonstration.Source: G. Discussion and Practical Implications · p. 11
  • A predicted regional field can provide location-specific generation context without requiring target stations to participate in its construction.The representation could also support regional monitoring, spatial anomaly detection, and distribution-level situational awareness.Source: G. Discussion and Practical Implications · p. 11
  • Evaluation covers a single regional fleet, and reconstructed fields depend on available source-station coverage and characteristics.The forecasting protocol assumes immediate satellite-observation availability, leaving operational product latency for real-time deployment unaddressed.Source: G. Discussion and Practical Implications · p. 11

IV. CONCLUSION

  • Across 496 systems, predicted future PV fields achieve regional field MAE of 0.0880 and reduce station-level MAE by approximately 10.0% versus the best single-station temporal baseline and 2.5% versus the matched graph-based pathway.The fleet comprises 446 source stations and 50 held-out target stations.Source: §IV · p. 11
  • The conclusion presents distributed PV fleets as regional sensing resources and spatial fields as a basis for richer spatiotemporal models of generation dynamics.The proposed pathway connects forecasted regional states with location-specific predictions.Source: §IV · p. 11
Loading 2610.16110v1…