Naslov
Very short-term prediction of wind farm power production with Deep Neural Networks

Autori
Đalto, Mladen ; Lončarek, Tomislav ; Vašak, Mario ; Matuško, Jadranko

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, pp prezentacija, znanstveni

Skup
2nd Frontiers in Computational Physics Conference: Energy Sciences

Mjesto i datum
Zürich, Švicarska, 03.06.2015. - 05.06.2015

Vrsta sudjelovanja
Poster

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
very short-term; wind farm power; deep neural networks

Sažetak
State of the art results achieved by deep learning methods in various fields suggest potential for improvements in very short-term wind power prediction. Performance of persistence based prediction and shallow neural networks used for wind power forecast have not yet been compared to deep learning methods. Use of large historical weather and SCADA datasets requires computational complexity reduction without sacrificing prediction quality. Issuing improved wind power forecasts for supporting decision-making in regulating reserve management has the advantage of being more cost-effective when compared to other solutions such as increasing backup capacities. Providing forecast uncertainty information further improves decision making capabilities and it is analysed in this paper. We present a comparison of simple persistence based prediction to prediction performance of shallow and deep neural networks. Input variable selection is performed in order to reduce complexity. Partial mutual information and compressing autoencoder approaches are compared. Proposed prediction techniques are tested on the case of on-shore wind farm Danilo in Croatia.

Izvorni jezik
Engleski

Znanstvena područja
Elektrotehnika, Temeljne tehničke znanosti, Informacijske i komunikacijske znanosti

POVEZANOST RADA