Robust estimation of wind power ramp events with reservoir computing

• Published in: Renewable energy Vol. 111; pp. 428 - 437
• Main Authors: Dorado-Moreno, M., Cornejo-Bueno, L., Gutiérrez, P.A., Prieto, L., Hervás-Martínez, C., Salcedo-Sanz, S.
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.10.2017
• Subjects: Echo state networks; Reanalysis data; Recurrent neural networks; Reservoir computing; Time series; Wind power ramp events prediction; Reanalysis data; Recurrent neural networks; Reservoir computing; Wind power ramp events prediction; Echo state networks; Time series
• ISSN: 0960-1481; 1879-0682
• DOI: 10.1016/j.renene.2017.04.016

Wind power ramp events are sudden increases or decreases of wind speed within a short period of time. Their prediction is nowadays one of the most important research trends in wind energy production because they can potentially damage wind turbines, causing an increase in wind farms management costs. In this paper, 6-h and 24-h binary (ramp/non-ramp) prediction based on reservoir computing methodology is proposed. This forecasting may be used to avoid damages in the turbines. Reservoir computing models are used because they are able to exploit the temporal structure of data. We focus on echo state networks, which are one of the most successfully applied reservoir computing models. The variables considered include past values of the ramp function and a set of meteorological variables, obtained from reanalysis data. Simulations of the system are performed in data from three wind farms located in Spain. The results show that our algorithm proposal is able to correctly predict about 60% of ramp events in both 6-h and 24-h prediction cases and 75% of the non-ramp events in the next 24-h case. These results are compared against state of the art models, obtaining in all cases significant improvements in favour of the proposed algorithm. •A direct binary classification approach for wind power ramp events is proposed, based on Reservoir Computing techniques.•Several novel ESN architectures are proposed, exploring different strategies for feeding the reservoir of the system.•Experiments in real data from three wind farms located in Spain are carried out.•Significant performance improvements are obtained by using the three architectures proposed.