Short-term wind power forecasting using simple recurrent genetically optimized neural networks.

Laburpena

The high penetration of wind power in electric power systems has become a relatively important matter that can affect daily power systems operation. The intermittent nature of wind makes it difficult to forecast electric power generated in wind farms over the following hours, and this may make it necessary to increase the power plant spinning reserve. This paper presents a new short-term wind power forecasting model. It is based on simple recurrent neural networks that enable the extraction of useful temporal information from input data thanks to their memorization capability. The recurrent neural network described in this paper was genetically optimized, with a novel structure for encoding all the parameters that define the neural network into specific numerical strings. The optimization process selects the inputs used for the neural network from a set of available inputs, the number of neurons in the hidden layers, the optimal values for the parameters of the training process and the value that controls the memorization capability of the neural network. The forecasting results obtained with this new model showed an improvement over the results obtained from a set of wind power forecasting models using the same input data.