To accurately predict the wind power and adopt methods to balance the fluctuation of power grid, an improved long short-term memory (LSTM) neural network wind power forecast algorithm based on noise reduction by threshold empirical modal decomposition (TEMD) is proposed. First, the actual operation and maintenance data of wind farms are normalized and divided into a training set and a test set. Then, an LSTM structure is designed and a Sub-Grid Search (SGS) algorithm is proposed to optimize the hyperparameters of the LSTM network. Finally, the power data are decomposed and noise-reduced using TEMD is improved by the variable-point technique and the TEMD-LSTM power forecast model is constructed to predict the power in time. The predicted values obtained are restored and evaluated by the original size. The results show that compared with five other algorithms of the same kind, the proposed algorithm in this paper has a root mean square error (RMSE) of 30.40, a trend accuracy (TA) value of 67.23% and a training time of 886 s, with the best overall performance.

为准确预测风电功率并采取平衡电网波动的方法，提出了一种改进的基于阈值经验模态分解（TEMD）降噪的长短期记忆（LSTM）神经网络风电功率预测算法。首先，将风电场实际运维数据归一化，分为训练集和测试集。然后，设计了 LSTM 结构并提出了子网格搜索 (SGS) 算法来优化 LSTM 网络的超参数。最后，对功率数据进行分解，并通过变点技术改进TEMD降噪，构建TEMD-LSTM功率预测模型，对功率进行及时预测。得到的预测值按原始大小进行恢复和评估。