Robust short-term prediction of wind turbine power based on combined neural networks

This paper proposes a powerful and accurate prediction method for modeling wind turbine output power considering uncertainty effects. It develops a probabilistic prediction method to make intervals around the forecast point samples. Due to the instable nature of the neural networks, several models with different classes are first trained, and then a combinatorial interval is constructed based on all their results. Through this idea, some lower and upper bounds are produced which can capture the uncertainty effects. In order to train the neural networks and adjust their setting parameters, a new optimization method based on the social spider optimization search (SOS) algorithm is presented which is equipped with some modification methods for increasing its capabilities. This increases the training robustness, thus resulting in stable but accurate algorithms. The experimental results of some wind turbines are used to assess the performance of the proposed method.