Abstract
Wind power forecasts of the expected wind feed-in for the next hours or days are necessary to integrate the generated volatile wind energy into power systems. Most forecasting models predict in some sense the best value, but they ignore the other possible outcomes which may arise because of forecasting uncertainties. Probabilistic forecasts, on the other hand, also predict a distribution of possible outcomes with their respective probabilities that specific power values will occur and therefore have higher information content. In this work, we address two problems that hinder the introduction of probabilistic forecasts in practice: (1) no measurement data are available for some wind farms and (2) the flexible aggregation of probabilistic forecasts for changing wind farm portfolios. We present an approach based on copulas that can solve both problems by modeling the spatial correlation structure between reference wind farms. By sampling from the resulting joint probability distribution, probability forecasts can be upscaled from reference wind farms to wind farms without power measurements. Furthermore, the results can be aggregated to probability forecasts of portfolios of arbitrary and changing size.
Introduction
As wind energy becomes a cornerstone of the power grid, traditional "point forecasts"—which predict a single expected value—are increasingly insufficient. Grid operators require probabilistic forecasts that quantify uncertainty to manage risks such as sudden supply drops. However, two major hurdles persist in the practical application of these forecasts. First, the "Data Gap" or upscaling problem: many wind farms lack real-time power measurement sensors, preventing traditional statistical models from being trained. Second, the "Portfolio Problem" or aggregation: simply summing individual probabilistic forecasts is mathematically incorrect due to spatial correlations between wind farms. Furthermore, portfolios are often dynamic, making fixed aggregate models difficult to maintain.
Methods
The proposed framework utilizes a multi-step pipeline combining local forecasting, spatial dependency modeling, and distance-based extrapolation.
The process begins with Local Probabilistic Forecasting for "reference" farms (those with available data) using Quantile Regression Neural Networks to predict 24 different quantiles of power output based on weather model data. The core innovation lies in Modeling Dependencies with Copulas. Gaussian Copulas are employed to model the spatial correlation structure between reference farms, transforming power variables into a joint normal distribution to capture how regional fluctuations move together.
From this joint distribution, the system performs Scenario Generation, drawing 1,000 samples that represent physically plausible snapshots of power output across the region. For "target" farms without data, an Upscaling step uses a distance-dependent weighting scheme to extrapolate scenarios from reference farms. Finally, Calibration via Ensemble Copula Coupling ensures predicted probabilities match observed frequencies, and Flexible Aggregation is achieved by simply summing the scenarios for the relevant farms in a portfolio.
Results
The study evaluated the model using a dataset of 20 wind farms in Northern Germany over a two-year period (2008–2009). The performance was measured using the Continuous Ranked Probability Skill Score (CRPSS) and reliability diagrams.
Key findings include:
- Individual Farm Accuracy: For target farms with no local data, the upscaled and calibrated model achieved a CRPSS of approximately 0.4, matching the performance of state-of-the-art models that had access to local data.
- Portfolio Reliability: For aggregate forecasts, the model achieved a CRPSS of 0.561.
- Calibration Impact: Calibration significantly improved reliability by widening distribution margins to correctly account for the increased uncertainty inherent in sites without local sensors.
The use of scenarios as a "universal currency" allows for the bypass of complex mathematical summation of probability distributions, making the approach highly scalable for dynamic portfolios where farms are added or removed frequently.
Conclusion
This work solves two critical problems hindering the introduction of quantile probabilistic forecasts in practical applications: the lack of measurement data at specific sites and the need for flexible portfolio aggregation. By leveraging spatial correlation through copulas, the method provides a pathway for advanced applications such as predictive grid security with stochastic optimization and improved marketing of renewable energies. Future work will focus on integrating temporal dependencies for short-term "nowcasting" and testing the model's robustness against real-world sensor noise and maintenance downtime.
Summary updated