Novel training strategies for wavelet-neuro models for wind speed prediction

The wind energy provides opportunities to generate power cheaply and cleanly without affecting the environment. The problem with wind energy is its variable and intermittent nature. Thus a large-scale introduction of wind power causes a number of challenges for the electricity market and power system operators who need to deal with the variability and uncertainty in wind power generation when making their scheduling and dispatch decisions. Accurate wind power forecast can solve these problems to a great extent. This paper proposes three novel strategies to train neural network to improve the prediction accuracy. Wavelet decomposition is used to filter out the high frequency outliers in the wind speed, thus making a smooth data to make the prediction accurate. The filtered data is used to train the neural network. In recursive training, the number of prediction steps during the training process, are reduced to increase the prediction accuracy. The neural network is re-trained with these predicted values. In conditional training, a pre-determined threshold level is set for the error. The training stops when the error falls below this level. In parallel training, 10 parallel networks is created with either recursive or conditional training, each of which is trained separately and the final predicted wind speed is the mean of the prediction done by individual parallel path.