PaperlayerView exact revision on arXiv

Source-linked AI summary

A TCN-based Hybrid Forecasting Framework for Hours-ahead Utility-scale PV Forecasting

Yiyan Li, Lidong Song, Si Zhang, Laura Kraus, Taylor Adcox, Roger Willardson, Abhishek Komandur, Ning Lu

arXiv:2111.08809v2eess.SP

TL;DR

Hours-ahead utility-scale PV forecasting must represent both stable hourly trends and rapid cloud-driven fluctuations. The paper develops a TCN-based hybrid framework that blends NWP inputs, forecasts cloud events from selected detector sites, and reconciles both outputs; its reported six-hour-ahead performance improves by approximately 20% over state-of-the-art methods.

  • Problem

    Physics-based models capture longer-horizon hourly trends but miss intra-hour cloud fluctuations, while data-driven models capture fluctuations but lose forecasting accuracy and stability with horizon.

  • Method

    The framework uses three TCNs for NWP blending, detector-based spatial-temporal cloud-event forecasting, and reconciliation of TF and CF outputs, with automated scenario-based detector-site selection.

  • Results

    The framework improves six-hour-ahead PV power forecasting performance by approximately 20% compared with state-of-the-art methods.

  • Takeaways & Limitations

    Reconciled forecasts preserve both long-term trends and intra-hour fluctuations by combining the complementary TF and CF forecasts.

Abstract

from arXiv · show

This paper presents a Temporal Convolutional Network (TCN) based hybrid PV forecasting framework for enhancing hours-ahead utility-scale PV forecasting. The hybrid framework consists of two forecasting models: a physics-based trend forecasting (TF) model and a data-driven cloud-event forecasting (CF) model. Three TCNs are integrated in the framework for: i) blending the inputs from different Numerical Weather Prediction sources for the TF model to achieve superior performance on forecasting hourly PV profiles, ii) capturing spatial-temporal correlations between detector sites and the target site in the CF model to achieve more accurate forecast of intra-hour PV power drops, and iii) reconciling TF and CF results to obtain coherent hours-ahead PV forecast with both hourly trends and intra-hour fluctuations well preserved. To automatically identify the most contributive neighboring sites for forming a detector network, a scenario-based correlation analysis method is developed, which significantly improves the capability of the CF model on capturing large power fluctuations caused by cloud movements. The framework is developed, tested, and validated using actual PV data collected from 95 PV farms in North Carolina. Simulation results show that the performance of 6 hours ahead PV power forecasting is improved by approximately 20% compared with state-of-the-art methods.

I. INTRODUCTION

The paper addresses complementary weaknesses in physics-based and data-driven PV forecasting by developing a TCN-based hybrid framework with automated detector-site selection. It combines stable hourly trend prediction with intra-hour cloud-event forecasting for hours-ahead utility-scale PV forecasting.

  • Cloud-movement stochasticity is a major cause of large PV forecasting errors, especially over hours-ahead horizons.
  • Existing methods lack a deep fusion approach that seamlessly integrates physics-based and data-driven forecasting models.
  • Physics-based models support longer-horizon hourly forecasting but struggle with intra-hour cloud-driven fluctuations, whereas data-driven models capture fluctuations but lose accuracy and stability with horizon.
  • Existing detector-site selection methods overlook cloud-event sequence, multi-site collaboration, and automated objective selection.
  • The proposed framework uses three TCNs for NWP blending, detector-based intra-hour cloud-event forecasting, and temporal reconciliation of trend and fluctuation forecasts.
  • A scenario-based automated algorithm identifies contributive neighboring sites without domain expertise or human supervision.

II. METHODOLOGY

The methodology combines physics-based trend forecasting and data-driven cloud-event forecasting through three TCN components. The resulting forecast is reconciled at 5-minute resolution to preserve both hourly trends and intra-hour fluctuations.

  • The TF model predicts hourly PV outputs for up to one week, while the CF model predicts 5-minute intra-hour fluctuations for the next 1 to 6 hours.
  • TCN #1 blends hourly NWP inputs from different sources before converting predicted irradiance into hourly PV forecasts with a physics-based model.
  • The CF model uses detector-network irradiance histories and TCN #2 to extract spatial-temporal correlations for intra-hour PV forecasting.
  • A. Temporal Convolutional Network: TCN uses zero padding, dilated convolutions, causal convolutions, and residual connections to extract temporal features without future-information leakage.
  • A. Temporal Convolutional Network: Dilated convolutions expand the receptive field through filter size, dilation rate, and network depth while limiting model complexity.

B. TCN-based NWP Data Blending

Different NWP sources have distinct forecasting features and granularity, creating discrepancies in physics-based trend forecasting. A TCN blender fuses these sources into a field-measurement irradiance sequence to improve the input to PV forecasting.

  • Different NWP sources vary in forecasting features and data granularity, producing discrepancies when used as physics-based model inputs.
  • The TCN-based blender performs sequence-to-sequence mapping from multiple NWP time series to field-measured irradiance.

C. Scenario-based Detector Sites Selection Algorithm

The detector-selection algorithm identifies neighboring sites whose cloud events lead those at the target site within the forecasting horizon. It uses event-based, time-lagged correlation and greedy network refinement before spatial-temporal TCN forecasting.

  • The algorithm targets neighboring sites that enable more accurate prediction of large PV drops caused by cloud movements.
  • A cloud event is an irradiance drop exceeding threshold ∆x across two consecutive intervals; here, ∆x = 0.3 captures drops above 30% of rated power in 5 minutes.
  • Time-lagged correlation finds ∆tmax, the shift maximizing Pearson correlation between detector and target cloud-event series.
  • Only leading correlations with ∆tmax > 0 can foresee target-site cloud events; high but lagging correlation cannot provide that forecasting information.
  • Successful detection requires a detector event to precede the target event by ∆tmax ∈ (0, Tthre], where Tthre equals the CF forecasting horizon.
  • The detector network is formed by adding neighbors in descending yearly Pcc.max order, retaining the subset with maximum φ, then removing sites if refinement improves φ.
  • The selected detector network supplies historical target and neighbor data to TCN #2 for extracting spatial-temporal correlations.

D. TCN-based Forecasting results reconciliation

Trend and cloud-event forecasts can disagree in temporal resolution and magnitude, so the framework reconciles them with a sequence-to-sequence TCN. The output preserves consistent hourly and 5-minute PV forecasts.

  • TF and CF forecasts require reconciliation because their separately produced profiles can differ in resolution and magnitude.
  • The reconciliation TCN aligns each hourly TF point with 12 five-minute CF points and trains against the target farm’s actual five-minute profile.
  • The trained model produces consistent five-minute and hourly PV outputs from the reconciled forecasts.

III. CASE STUDY

The case study uses five-minute measurements from 95 utility-scale PV farms in North Carolina to develop and verify the proposed algorithm. The farms span 0.4 MW to 26.2 MW, with pyranometer irradiance used to derive cloud movements.

  • The study uses five-minute field data from 95 utility-scale PV farms in North Carolina collected from 1/1/2020 to 11/30/2020.
  • PV farm sizes range from 0.4 MW to 26.2 MW, with locations documented for the case study.
  • Pyranometer irradiance measurements are selected to derive cloud movements because horizontal placement makes measurements more comparable across sites.

A. TF model: TCN-based NWP Data Blending

The study evaluates TCN-based forecasting components and detector-selection strategies for utility-scale PV forecasting. Results show that TCN blending improves trend forecasting, while selected detector sites improve cloud-event forecasting accuracy and stability.

  • TCN blending reduces TF forecasting bias to 0.47 and RMSE to 43.17, outperforming unblended NWP inputs and five benchmark blending methods.
  • Using only selected neighbors gives all evaluated models the smallest RMSE median and IQR, demonstrating the efficacy of the neighbor-selection method.
  • Using all sites produces large RMSE variances, while target-site or randomly selected neighbors produce larger errors.
  • TCN with detector sites achieves the lowest RMSE, best forecasting stability, and best computation efficiency.
  • Detector sites mainly improve 2–3-hour-ahead PV forecasting, while errors increase dramatically after hour 4 and exceed those of the physics-based model.

C. TCN-based Forecast Reconciliation

Forecast reconciliation combines the complementary strengths of CF and TF forecasts. It preserves intra-hour fluctuations while restoring longer-term trends, reducing errors at longer forecasting horizons.

  • CF forecasts deviate from actual power beyond 2 hours because real-time inputs do not capture long-term trends, whereas NWP-based TF forecasts capture trends but miss intra-hour fluctuations.
  • Reconciliation combines independently generated TF and CF time series to preserve both intra-hour fluctuations and long-term trends.
  • Across all 95 PV sites, reconciliation significantly reduces forecasting RMSE for horizons longer than 2 hours.
  • Figures 12 and 13 compare CF forecasting examples and RMSE before and after reconciliation.

IV. CONCLUSION

The paper concludes that its TCN-based hybrid framework improves hours-ahead utility-scale PV forecasting by combining physics-based trend prediction, data-driven cloud-event prediction, and forecast reconciliation.

  • TCN improves NWP blending in TF, spatial-temporal forecasting in CF, and reconciliation of forecasting results.
  • TCN-based NWP blending improves forecasting accuracy by 37%.
  • Forecast reconciliation provides another 26% performance improvement by merging TF and CF forecasting results.
  • The framework uses selected neighboring sites to capture spatial-temporal correlations associated with large PV output drops.
  • The authors characterize TCN as a comprehensive data-driven solution for short-term PV forecasting with superior accuracy and computation efficiency.

V. APPENDIX

The appendix documents the three TCN configurations and examines how receptive-field hyperparameters affect performance. Performance remains stable when the receptive field covers the full input sequence but degrades when it does not.

  • The framework uses three TCNs for NWP blending, spatial-temporal forecasting, and forecasting-results reconciliation.
  • TCN hyperparameters are selected so the receptive field covers the complete input sequence.
  • Performance is stable across kernel-size, dilation-rate, and stack-number combinations when the receptive field is sufficient.
  • Model performance degrades when the receptive field is insufficient to cover the input sequence.
  • The appendix includes configurations for the TCN models and the VGG-8 benchmark network.

C. CNN-LSTM

CNN-LSTM and GARNN are benchmark deep-learning models for spatial-temporal PV forecasting. CNN-LSTM has approximately 66K capacity, while GARNN models dynamic inter-site correlations and has approximately 64K capacity.

  • C. CNN-LSTM: CNN-LSTM is presented as a popular deep-learning model for spatial-temporal PV forecasting.The paper's benchmarking CNN-LSTM model has approximately 66K capacity.
  • C. CNN-LSTM: The paper provides configuration tables for both the benchmarking CNN-LSTM and GARNN models.These are identified as Table XI and Table XII, respectively.
  • C. CNN-LSTM: GARNN represents PV sites as a graph using an adjacency matrix that can vary with dynamic weather conditions.A multi-head attention mechanism learns the dynamic adjacency matrix from each site's time-series data.
  • C. CNN-LSTM: GARNN feeds the learned dynamic matrix and original time-series data into an RNN to extract temporal information for forecasting.The paper reports GARNN model capacity of approximately 64K.
Loading 2111.08809v2…