Speaker
Description
Accurate forecasts of regional wind power production are crucial for power system operation and planning. Total generation in the coming hours and days depends strongly on both weather forecasts and time-varying production capacity. We propose a probabilistic forecasting approach based on Bernstein Quantile Networks (BQN) to predict aggregated power production from multiple wind farms without access to individual farm data.
The model uses ensemble forecasts of several meteorological variables at each wind farm, combined with time-dependent capacity information. Its architecture exploits the structured input through (i) learnable wind farm embeddings, (ii) mappings to power output, (iii) spatial aggregation across farms, and (iv) permutation-invariant layers capturing distributional features. The BQN framework produces full predictive distributions via quantile functions represented by Bernstein polynomials.
