Wind speed interval prediction based on multidimensional time series of Convolutional Neural Networks

Abstract

Wind power, as an economical and clean energy source, has rapidly infiltrated into modern power grids. Wind speed prediction is a pivotal technology for wind power integration, which has attracted much attention from researchers and practitioners. Nevertheless, traditional methods only focus on point prediction, which is far from meeting the requirements of power system risk assessment and risk control. To fill this gap, we propose an interval prediction and analysis system based on data preprocessing, deep learning, convolutional neural networks, and interval prediction. It can effectively predict and analyze the uncertainty in short-term wind speeds. First, an original wind speed sequence is divided into high-frequency and low-frequency data using a data preprocessing strategy. After deleting the high-frequency noise data, the denoised multidimensional time series is entered as input into an optimized convolutional neural network. Subsequently, the lower and upper bounds of the interval prediction are obtained using the interval construction principle and variance of the training data set. To verify the effectiveness of the proposed hybrid system, a 10 min wind speed data set of the Shandong wind farm in China is used as an example for analysis. The experimental results show that the proposed hybrid prediction system can not only effectively predict the variation trend of wind speed series, but also accurately predict the local characteristics and volatility of wind speed with strong randomness.