An ultra-short-term wind power prediction method based on CEEMDAN, WPE, CNN, LSTM and SAM is proposed. Firstly, the original wind power time series is decomposed by the CEEMDAN to reduce the nonlinearity and volatility of the original series. Secondly, the similarity between components is calculated by WPE method, and the similar components are recombined to correct the over decomposition problem of CEEM-DAN and make the modified modal components more regular. Finally, the recombined components are input into the CNN-LSTM network for time series modeling, and the neural weights of this network are redistributed by SAM, which improves the adaptability of the network to the uncertainty of input characteristics. On this ba-sis, the mechanism of SAM, CEEMDAN and WPE in wind power prediction and the crucial physical informa-tion contained in wind power signals are clarified, which proves the effectiveness of them in wind power signal modal decomposition and the hidden layer output weight allocation of LSTM.
关键词
超短期风电功率预测/自适应噪声完全集合经验模态分解/加权排列熵/卷积长短期记忆网络/自注意力机制
Key words
ultra-short-term wind power prediction/complete ensemble empirical mode decomposition with adaptive noise/weighted permutation entropy/convolutional long-short-term memory network/self-attention mechanism
维普
万方数据