A novel approach to multi-horizon wind power forecasting based on deep neural architecture. (November 2021)

• Record Type: Journal Article
• Authors: Putz, Dominik; Gumhalter, Michael; Auer, Hans
• Is Part Of: Renewable energy. Volume 178(2021)
• Page Start: 494
• Page End: 505
• Publication Date: 2021-11
• Subjects: Wind power forecasting -- Neural networks -- Deep learning -- N-BEATS -- Pinball-sMAPE; Renewable energy sources -- Periodicals; Power resources -- Periodicals; Énergies renouvelables -- Périodiques; Ressources énergétiques -- Périodiques; 333.794
• DOI: 10.1016/j.renene.2021.06.099
• Languages: English
• ISSNs: 0960-1481

Abstract

Abstract: In recent years, renewable energy sources have been installed in large numbers. Wind power in particular, a technology with very high potential, has become a significant source of energy in most power grids. However, wind power generation forecasting and scheduling remain very difficult tasks due to the uncertainty and stochastic behaviour of wind speed. This work provides a novel, powerful tool for wind power forecasting based on neural expansion analysis for time series forecasting (N-BEATS), a deep neural network approach. N-BEATS was designed as an easy-to-implement approach to solving non-linear stochastic time series forecasting problems. Additionally, a loss function is tailored to wind power forecasting to confront the issue of forecast bias. The results are compared with established models, such as statistical and machine learning approaches as well as hybrid models, using the real-world wind power data from 15 different European countries as input. Comprehensive and accurate results are obtained during this work, showing that this methodology can easily compete with other approaches and even outperform them in terms of accuracy in most cases. Additionally, the tailored loss function reduces the error significantly. The N-BEATS architecture is further customized to deliver decomposed components such as trend and seasonality, yielding interpretable outputs. These findings constitute considerable progress and provide support for decision makers. Highlights: Deep neural architecture for short-term wind power production. Forecast bias addressed by a tailored loss function. Decomposed and interpretable results to facilitate decision making.