Posters
Come meet the poster presenters to ask them questions and discuss their work
We would like to invite you to come and see the posters at our upcoming conference. The posters will showcase a diverse range of research topics, and will give delegates an opportunity to engage with the authors and learn more about their work. Whether you are a seasoned researcher or simply curious about the latest developments in your field, we believe that the posters will offer something of interest to everyone. So please join us at the conference and take advantage of this opportunity to learn and engage with your peers in industry and the academic community.
PO382: Combining Deep Learning, PCA and Gradient Boosting for Short-Term Wind Energy Forecasting
Ignacio Villanueva, CEO and Full Professor in Mathematics, Ravenwits
Abstract
We present two new methodologies for short-term wind energy forecasting (up to 100 hours). Both approaches focus on feature reduction—one through intelligent extraction with Convolutional Neural Networks (CNNs), and the other through a more naive dimensionality reduction using PCA—followed in both cases by a gradient boosting model performing the regression task. These methods build on our previous work in wind energy forecasting with CNNs. Our CNN forecasters, developed in collaboration with the Spanish TSO, Red Eléctrica (REE), outperformed all models currently used or purchased by REE for the Spanish grid. A main limitation of Deep Learning (DL) in this context stems from the fact that actual wind conditions are unknown. Instead, we rely on forecasts that represent averages of the last ten minutes in each hour. Consequently, our inputs—the Numerical Weather Predictions (NWP)—contain significant noise, which limits CNN effectiveness. CNNs usually excel when trained on abundant, high-quality data, which is not the case here. Classical statistical methods, as well as Regression Trees, including Random Forest and Gradient Boosting (GB), are generally better suited for problems with lower-quality data. We propose a hybrid methodology that combines DL with these methods, together with a simplified feature extraction step using PCA. The first approach employs an encoder-decoder trained to reconstruct the newest forecast when given an older one for the same region and time. The latent space captures biases and complex weather patterns, approximating the representation of a shorter forecast horizon. These features are then used as input to a gradient boosting model. The second is an architecture that applies PCA to NWP data and then uses the resulting components as input to a tabular model as above.
No recording available for this poster.
