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

In recent years rapid growth of wind power generation in many countries around the world has highlighted the importance of wind prediction. In this work neural networks are used for ultra-short-term wind prediction. In many instances reported in the literature neural network exhibit poor performance- very often because no complexity reduction methods were considered. To that end, in this paper two input variable selection algo-rithms based on partial mutual information are compa-red for further use with nonlinear models such as neural networks. Performance improvements of the proposed prediction system are compared to neural networks wit-hout input variable selection, and validated for locations near Split, Croatia. The use of neural network drastically outperforms simple persistence estimator on 3 hour horizon. 1