Improvements in wind speed forecasting using an online learning

The use of wind energy has developed significantly worldwide. Wind power is the strongest growing form of renewable energy, ideal for a future with pollution-free electric power. But the intermittent nature of wind makes it difficult to forecast. The researchers have embarked on use of the Multi-Layered Perceptrons (MLP) neural networks and other architectures of neural networks for predicting the actual wind speed from the previous values of the same variable. Several algorithms are then supplied with the data to establish the relationship between the inputs and the output. The obtained results indicate that the identified model can successfully be used, but with a poor accuracy prediction of the wind speed and the Forecasting error increases as we go far dates (hours or minutes) of learning data. In this work a new neural networks approach is developed for predicting the actual wind speed from the previous values of the same variable. Model parameters are estimated from a set of past available data, and they are regularly updated during online operation by accounting for any newly available information. The learning involves physical changes of the connections between neurons. The association between several neural structures, with a specific function, allows the emergence of a higher-order function for all. If the neural network can not monitor the physical changes over time we opt for a dynamic learning over time. Finally, we present a mathematical foundation for our approach.