【摘要】 为研究单台风机数据进行风功率预测时未考虑空间特征造成预测精度不理想的问题，提出了一种考虑多对一时空特征的基于改进序列到序列(Seq2Seq)模型的短期风功率预测组合模型。首先，采用k近邻算法对风电场的风机实现空间区域的划分，获取k台近邻风机的数据，基于孤立森林算法对异常数据进行识别、筛选和填充。其次，应用双向门控循环单元和自注意力机制对Seq2Seq模型进行改进，利用具有空间特征的邻接矩阵对模型进行权重优化。最后，进行多对一短期风功率预测，输出目标风机风功率预测结果。采用美国风场实际运行数据，将所提出的组合模型与长短期记忆(LSTM)等5种模型进行对比，以验证模型的可靠性。结果表明：该组合模型在时空风功率预测中表现出令人满意的稳定性和鲁棒性，可有效提高风功率预测精度及效率。更多还原

【Abstract】 Since the prediction accuracy of wind power from a single wind turbine was not so satisfactory without the consideration of spatial-temporal features, a many-to-one combined model for short-term wind power prediction has been proposed based on a modified sequence to sequence(Seq2Seq) model. Firstly, the k-nearest neighbor(k-NN) algorithm was used to divide the wind turbines from wind farm into spatial regions and obtain the data of k neighboring wind turbines. The abnormal data were identified, filtered and filled by isolation forest(IF) algorithm. Then the bi-directional gated recurrent unit(BiGRU) and self-attention(SA) mechanism were adopted to improve the Seq2Seq model, and the adjacency matrix with spatial features was used for weight optimization in the model. The many-to-one short-term wind power prediction was performed along with the wind power output of target wind turbines. The proposed combined model was compared with five other models such as long short-term memory(LSTM) to verify its reliability with the actual operation data of wind farms in the United States. Results show that the combined model shows satisfactory stability and robustness in spatial-temporal wind power prediction and effectively improves its accuracy.更多还原