An improved convolutional neural network-based approach for short-term wind speed forecast

Wind power is taking more and more important proportion in new energy power system, where wind speed forecast (WSF) plays a key role in maintaining safe and smooth operation of wind power generation system. However, with wind speed influenced by complex meteorological and topographic factors, the current models based on raw historical data have encountered many problems. Therefore, the accurate WSF has been a great challenge in practice. To solve this problem, a novel WSF method based on convolutional neural network (CNN) is proposed in this paper. By utilizing the big data collected from the running log of a real wind farm, the convolution layers of CNN can automatically extract the deep features from the historical data. The characteristics of sparse connection and weight sharing contributes to a faster training speed of CNN. In order to find the optimal CNN structure suitable for WSF, the parameters are determined one by one through experimental analysis. The CNN model is compared with the traditional multi-layer perception (MLP) network to show the advantages in WSF. In the full connection layer, the improved Biased Dropout method is applied to effectively reduce the number of parameters and reduce the complexity of the model. The performance of the modified CNN is evaluated through a WSF test in every ten minutes, which verifies that the training speed is accelerated in comparison with that of the typical CNN.