To improve the accuracy of multi-step wind power forecast, a variational mode decomposition-long short-term memory (VMD-LSTM) forecast method is proposed. Firstly, the variational mode decomposition method is adopted to decompose the wind power data into three constituent modes, named as the long-term component, the fluctuation component and the random component. Secondly, long short-term memory network is utilised to deeply learn the characteristics of the three constituent modes. Profit from its unique forget gate and memory gate structure, the association with long-term time series is learned to build a multi-step forecast model. Finally, the wind power data from ELIA and NERL are used to test. The error analysis shows that the proposed method has superior performance in the multi-step forecast and real-time forecast.

精准的电力负荷预测有利于保障电力系统的安全、经济运行。针对现行预测算法存在的预测准确度低、模型耗时长等问题，提出一种基于随机森林(random forest，RF)算法和粗糙集理论(rough set theory，RST)的改进型深度学习(deep learning，DL)短期负荷预测模型(RF-DL-RST)。该模型首先基于历史数据，利用随机森林算法提取影响负荷预测的关键特征量；然后将关键特征量和历史负荷值作为深度神经网络的输入、输出项进行训练，并通过粗糙集理论修正预测结果。最后，通过算例进行仿真验证，结果表明，该模型的预测准确度比单一的深度学习模型及不进行预测修正的模型更高。

Accurate power load forecasting is conducive to ensuring the safe and economic operation of the power system. Aiming at the problems of low prediction accuracy and long time consuming of the current prediction algorithms, an improved deep learning (DL) short-term load forecasting model based on random forest (RF) algorithm and rough set theory (RST), namely RF-DL-RST, was proposed. Firstly, based on historical data, the model used RF algorithm to extract the key features that affected the load forecasting. Then, the key features and historical load data were trained as the input and output items of deep neural network (DNN), and the prediction results were corrected by RST. After that, the rough set method was used to revise the prediction results. Finally, the simulation was verified by an example. The results show that the prediction accuracy of the model is higher than that of a single DNN model and a model without RST revised.

针对风电数据波动性过大而导致的风电功率预测不精确问题，提出一种基于自适应噪声完备集合经验模态分解(complementary ensemble empirical mode decomposition with adaptive noise,CEEMDAN)与蜣螂算法(dung beetle optimizer,DBO)优化的广义回归神经网络(generalized regression neural network,GRNN)超短期风电功率预测方法。首先将原始风功率序列进行时滞特性分析，选取与预测时刻相关性强的时序进行多路时序建模；然后对相关性强的时序进行CEEMDAN分解，得到一组本征模态分量(intrinsic mode function,IMF)和剩余分量；其次将上述两组分量输入经蜣螂优化算法优化的GRNN网络进行各分量预测；然后将各预测分量叠加，得到最终预测结果。算例分析表明，所提的CEEMDAN-DBO-GRNN预测模型的预测精度更高，而且CEEMDAN能够减少风电功率波动性与随机性对预测结果的影响，同时利用蜣螂算法优化后的超参数模型进行预测，在一定程度上提高了超短期风电功率预测的精度。

To address the problem of inaccurate wind power prediction caused by the excessive volatility of wind power data, this paper proposes a generalized regression neural network (GRNN) method based on the optimization of complementary ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and dung beetle optimizer (DBO). A combination of GRNN and DBO optimization is used for ultra-short-term wind power prediction. First, the original wind power sequence is subjected to time-lag characteristic analysis, and the time series with a strong correlation with the predicted moments is selected for multiplexed time-series modeling. Subsequently, the time series with strong time series are subjected to CEEMDAN decomposition, and a set of intrinsic mode functions (IMFs) and a residual term are obtained. Second, the two sets of the above components are inputted into the GRNN network optimized by the DBO algorithm for the prediction of the components. Subsequently, the prediction components are superimposed to obtain the final prediction result. Example analysis shows that the CEEMDAN-DBO-GRNN prediction model proposed in this paper has higher prediction accuracy, and CEEMDAN can reduce the influence of wind power volatility and randomness on the prediction results. The prediction of the hyperparameter model optimized by the DBO algorithm improves the accuracy of the ultra-short-term wind power prediction to a certain extent.