Abstract: A high-precision combinational method is presented to forecast the total-power of wind farm directly. Taking into account that the Non-stationary identity of wind speed leads to the non-stationary time series of total-power, the NARX neural network is adopted as the original forecasting model. And then, a hybrid forecasting method based on Empirical Mode Decomposition（EMD） and NARX neural network is proposed to improve the forecast precision. First, the total-power time series is decomposed into several stable trend terms with different Intrinsic Mode Functions（IMF）. Subsequently, the corresponding prediction models of NARX neural network are set up according to different stable components. These forecasting results of each component model are summed up in equal weight to obtain the final predictive value. Besides, the data of time intervals of 5-minute and 15-minute obtained from a large wind power plant are used in the experiments to explore how time intervals affect the predictive results. The experiment shows that the combinational model is suitable for the prediction of total-power. According to the experimental results, the combinational model is more accurate than many traditional methods, and the prediction accuracy of 5-minute time interval data is more accurate than that of 15-minute time interval data.

Key words: empirical mode decomposition, Nonlinear Auto-Regressive with eXogenous input neural network（NARX）, non-stationary time series, wind power, total power

摘要： 探索构建对风电场总功率进行直接预测的高精度组合预测算法。考虑到风速的非平稳性导致风电总功率表现为非平稳时间序列，采用NARX神经网络作为初步预测模型，提出了经验模态分解与NARX神经网络相结合的混合预测模型。对风电场总功率非平稳时间序列进行经验模态分解，得到不同频带本征模式分量的平稳序列。对不同频带的平稳分量建立相应的NARX神经网络预测模型，并将各分量模型的预测值进行等权求和得到最终预测值。此外，为研究不同时间间隔对预测结果的影响，采用某大型风电场时间间隔为5 min与15 min的数据进行实验。预测结果表明，提出的组合预测模型适合于总功率预测，其预测效果比传统模型的效果更佳，且时间间隔为5 min的数据比时间间隔为15 min的数据预测精度更高。

关键词: 经验模态分解, 非线性自回归神经网络（带外部输入的）（NARX）, 非平稳时间序列, 风电场, 总功率