Back to the programme printer.gif Print




Delegates are invited to meet and discuss with the poster presenters in this topic directly after the session 'Forecasting: Maximizing grid deliverability and leading your business processes to profitability' taking place on Thursday, 13 March 2014 at 11:15-12:45. The meet-the-authors will take place in the poster area.

Mladen Đalto Faculty of Electrical Engineering and Computing University of Zagreb, Croatia
Co-authors:
Mladen Đalto (1) F P Mario Vašak (1) Mato Baotić (1) Jadranko Matuško (1) Kristian Horvath (2)
(1) Faculty of Electrical Engineering and Computing University of Zagreb, Zagreb, Croatia (2) Meteorological and Hydrological Service, Zagreb, Croatia

Printer friendly version: printer.gif Print

Presenter's biography

Biographies are supplied directly by presenters at EWEA 2014 and are published here unedited

Mladen Đalto has received his BSc and MSc degree in electrical engineering and information technology from University of Zagreb, Faculty of Electrical Engineering and Computing (UNIZG-FER) in 2011 and 2013, respectively. He is currently employed at the Department of Control and Computer Engineering at UNIZG-FER as a research assistant within Advanced Control Team. His main research interests are in the areas of intelligent control systems and optimal control.

Abstract

Neural-network-based ultra-short-term wind forecasting

Introduction

In recent years, environmental considerations have prompted the use of wind power as a renewable energy resource. Rapid growth of wind power generation in many countries around the world has highlighted the importance of wind prediction. The reasons are twofold: facilitating integration of wind energy in the electricity system and advanced control of wind turbines and farms. Croatian wind energy sector lacks efficient wind energy management, and in particular a dedicated state-of-the-art wind system and wind forecasting system designed for the specific and challenging wind climate in Croatia.

Approach

Numerical weather prediction lacks ultra-short-term prediction capabilities mainly because of the time-consuming simulations of atmospheric models, especially if data assimilation is a component of the prediction system. Besides providing improved forecast up to 3 hours ahead, potential applications of ultra-short-term wind prediction are in optimal wind turbine and wind farm control.
In this work neural networks are trained on historical measurement and historical forecast datasets and then used online for ultra-short-term wind forecasting. Large amount of redundant and non-informative data inhibits the training algorithm performance. To cope with this issue here an input variable selection based on partial mutual information is used.


Main body of abstract

Ultra-short term wind speed and direction prediction can be used for operation mode selection of the wind turbine control system, nacelle wind following strategy, wind farm control etc.
Several ultra-short-term wind speed and direction forecasting methods have been reported in the literature over the past years. Most of them are based on extraction of causal relations on large historical measurement datasets.

Frequently used time-series models with well-developed theoretical background lack structural capabilities for modelling complex dynamics such as that of the wind. Often neural networks used do not include appropriate complexity reduction methods which results in lower performance. Input variable selection based on partial mutual information is found appropriate for use with nonlinear models such as neural networks. Partial mutual information algorithms based on k nearest neighbours and probability density kernel estimation are compared.

Developed neural network methodology can be used efficiently at a low computation cost for any location with varying number of input variables and historical data samples. Performance improvements of the proposed prediction system relative to simple persistence and to commonly used neural prediction methods are evaluated for locations near Split, Croatia.


Conclusion

Developed neural network methodology can be efficiently used on any location with associated sufficiently informative measurement dataset. Wide application to wind prediction on various sites is possible due to neural network structural ability to handle varying number of input variables and model highly nonlinear input-output relationships. Partial mutual information method is successfully used for input variable selection which consequently reduces complexity and neural network training time.


Learning objectives
-assessment of ultra-short-term wind prediction performance, used for wind turbine and wind farm control purposes,
-assessment of performance limit of data-based modelling procedures in wind forecasting,
-adaptation of input variable selection methods for wind.