Adaptive spatiotemporal graph learning for multi-horizon probabilistic wind power forecasting | JoVE Visualize

Area of Science:

Renewable Energy Systems
Artificial Intelligence in Energy
Power Systems Engineering

Background:

Accurate wind power forecasting is vital for grid stability and economic efficiency in systems with high renewable energy penetration.
Existing deep learning methods often overlook spatiotemporal dependencies, uncertainty calibration, and temporal coherence, limiting real-world applicability.
Challenges include optimizing for isolated horizons/locations rather than holistic spatiotemporal performance.

Purpose of the Study:

To propose a novel adaptive spatiotemporal graph neural network (ST-GNN) framework for multi-horizon probabilistic wind power forecasting.
To jointly optimize deterministic accuracy, probabilistic uncertainty calibration, and temporal coherence in forecasts.
To capture dynamic inter-site correlations and improve decision-making in power system operations.

Main Methods:

Developed an adaptive ST-GNN integrating site-specific meteorological and SCADA data.
Employed a dynamically evolving graph structure with adaptively reweighted edges to model changing inter-site correlations.
Formulated a unified multi-objective loss function combining error minimization (RMSE), probabilistic metrics (CRPS), and temporal smoothness regularization.

Main Results:

Achieved up to 13% RMSE reduction at 1-4 hour horizons and over 5% at 24-hour horizons compared to state-of-the-art baselines.
Demonstrated reduced calibration error with empirical coverage deviations within 3% across central quantiles.
Showcased significantly higher temporal rank-order consistency, improving multi-site coordination and reducing dispatch instabilities.

Conclusions:

The proposed adaptive ST-GNN framework offers a flexible and potentially generalizable approach for accurate, reliable, and coherent wind power forecasting.
Integrating spatiotemporal graph structures with probabilistic training objectives effectively addresses limitations of existing forecasting methods.
The model's ability to adapt to different meteorological regimes and maintain temporal stability enhances its practical value for grid operations.