[Paper Review] Synthetic Photovoltaic and Wind Power Forecasting Data

Stephan Vogt, Jens Schreiber+1 more|arXiv (Cornell University)|Apr 1, 2022

Energy Load and Power Forecasting4 citations

TL;DR

This paper introduces a large-scale, publicly available synthetic dataset of photovoltaic and wind power forecasts for 120 PV and 273 wind power plants across Germany, generated using real weather measurements and physical models. The dataset enables realistic machine learning research in renewable energy forecasting, with results showing that model errors on synthetic data closely match real-world historical data, establishing a reliable benchmark for future studies.

ABSTRACT

Photovoltaic and wind power forecasts in power systems with a high share of renewable energy are essential in several applications. These include stable grid operation, profitable power trading, and forward-looking system planning. However, there is a lack of publicly available datasets for research on machine learning based prediction methods. This paper provides an openly accessible time series dataset with realistic synthetic power data. Other publicly and non-publicly available datasets often lack precise geographic coordinates, timestamps, or static power plant information, e.g., to protect business secrets. On the opposite, this dataset provides these. The dataset comprises 120 photovoltaic and 273 wind power plants with distinct sides all over Germany from 500 days in hourly resolution. This large number of available sides allows forecasting experiments to include spatial correlations and run experiments in transfer and multi-task learning. It includes side-specific, power source-dependent, non-synthetic input features from the ICON-EU weather model. A simulation of virtual power plants with physical models and actual meteorological measurements provides realistic synthetic power measurement time series. These time series correspond to the power output of virtual power plants at the location of the respective weather measurements. Since the synthetic time series are based exclusively on weather measurements, possible errors in the weather forecast are comparable to those in actual power data. In addition to the data description, we evaluate the quality of weather-prediction-based power forecasts by comparing simplified physical models and a machine learning model. This experiment shows that forecasts errors on the synthetic power data are comparable to real-world historical power measurements.

Motivation & Objective

• To address the lack of publicly available, comprehensive datasets for machine learning-based renewable energy forecasting.
• To provide a large-scale, realistic synthetic dataset with precise geographic coordinates, timestamps, and physical plant characteristics.
• To support advanced ML research in transfer learning, multi-task learning, and zero-shot learning by including detailed static metadata for each power plant.
• To establish a benchmark for forecast error by comparing physical models and gradient boosted regression trees (GBRT) on synthetic data.
• To ensure synthetic data reflects real-world forecast errors by basing it on actual numerical weather prediction (NWP) inputs.

Proposed method

• Synthetic power time series are generated using physical models that simulate power output based on real historical weather measurements from the Icosahedral Nonhydrostatic (ICON) model.
• The dataset includes 500 days of hourly-resolution data (Dec 8, 2018 – Jun 2, 2020) covering all four seasons, enabling robust model training and testing.
• Each power plant is assigned unique physical and geometric parameters, such as PV module tilt and orientation, and wind turbine rotor-generator ratios, to ensure realism and enable transfer learning.
• Non-synthetic input features—such as wind speed, global horizontal irradiance, and temperature—are derived from the ICON-NWP model and paired with synthetic power outputs.
• A gradient boosted regression tree (GBRT) model is trained and compared against physical baseline models (e.g., Enercon and McLean power curves) to evaluate forecast accuracy.
• The dataset supports evaluation under data-scarce conditions by enabling experiments with truncated training sets (e.g., 7 to 365 days).

Experimental results

Research questions

• RQ1 How do forecast errors of machine learning models on synthetic data compare to those on real-world historical data?
• RQ2 Can synthetic data support reliable benchmarking for renewable power forecasting models, particularly in transfer and multi-task learning?
• RQ3 What is the performance gap between physical models and machine learning models (e.g., GBRT) under varying amounts of training data?
• RQ4 How do different empirical power curves (e.g., McLean, Enercon) compare in accuracy when physical turbine parameters are unknown?
• RQ5 To what extent does the inclusion of detailed plant-specific metadata improve model generalization and transfer learning performance?

Key findings

• The forecast errors of both the GBRT and physical baseline models on the synthetic dataset are comparable to those observed in real-world historical power measurements.
• For photovoltaic forecasting, the GBRT model outperforms the physical baseline when sufficient training data (e.g., 365 days) is available, achieving a mean nRMSE of 0.072 compared to 0.085 for the physical model.
• For wind power forecasting, the GBRT model consistently outperforms the Enercon physical baseline, reducing mean nRMSE from 0.210 (baseline) to 0.125 with full training data.
• With limited training data (e.g., 7 days), the physical model remains more robust for photovoltaics, while the GBRT model shows superior performance for wind power even with as little as 14 days of data.
• The McLean empirical power curve provides a viable alternative baseline for wind forecasting when turbine-specific parameters are unavailable, with nRMSE values close to the Enercon baseline (0.212–0.239).
• The synthetic dataset enables reliable evaluation of transfer learning and zero-shot learning due to the inclusion of detailed plant-specific metadata such as orientation, rotor diameter, and hub height.