基于低频波动挖掘和高频校正的风电超短期预测

Ultra-short-term Prediction of Wind Power Based on Low-frequency Fluctuation Mining and High-frequency Correction

相较于风电高频分量较高的随机性,风电的低频波动分量更能代表风电的未来走向。为此,提出一种基于低频波动挖掘和高频校正的风电超短期预测方法。为解决在全部低频过程中利用输入风电序列直接挖掘相似过程时出现的模式不匹配的问题,采用分类挖掘的方式。首先将低频波动过程划分为上爬坡、下爬坡、低出力、恒出力和其他5类模式,根据历史风电数据分别获取5类模式集。然后考虑输入序列的波动变化趋势将其截取为目标序列,在判断目标序列所属类别的基础上,挖掘所属模式集中的相似波动过程,进而得到相似波动的未来趋势,将其组成未来趋势矩阵。为解决直接依据低频分量得到预测结果的高频分量缺失问题,以未来趋势矩阵为输入,以风电实际值为输出,建立基于长短期记忆(long short-term memory,LSTM)的高频校正模型,利用LSTM模型学习低频波动性输入与实际输出的关联性,弥补低频波动模式中的高频分量,使得模型预测结果既利用波动性挖掘保证模式相似的准确性,又学习了高频分量保证风电功率的完整性。最后,以美国可再生能源实验室提供的加州海岸某风电场数据进行算例分析,验证所提方法的可行性和有效性。

Compared with the higher randomness of the high-frequency components of the wind power,the low-frequency fluctuation components of the wind power can better represent the future trend of wind power.This paper proposes a wind power forecasting method based on the low-frequency fluctuation mining and the high-frequency correction.In order to solve the problem of pattern mismatching that occurs when the input wind power sequence is used to directly mine similar processes in all the low-frequency processes,this paper adopts a classification mining method.First,the low-frequency fluctuation process is divided into five patterns,and the five pattern sets are obtained based on the historical wind power data.Then the fluctuation trend of the input sequence is intercepted as the target sequence.After judging the type of the target sequence,the similar fluctuation process in the pattern set is mined,getting the future trend of the similar fluctuation and composing it into the future trend matrix.In order to solve the problem of the prediction results without high-frequency components based on the low-frequency components,taking the future trend matrix as the input and the wind power actual value as the output,a high-frequency correction model based on long/short-term memory(LSTM)is established.The LSTM model is used to learn the relevance of the low-frequency volatility input and the actual output compensating the high-frequency components,so that the prediction results not only ensures the accuracy of the model with the volatility mining,but also guarantees the integrity of the wind power with the high-frequency components.Finally,the data of a wind farm on the coast of California provided by the National Renewable Energy Laboratory,U.S.A is used to conduct a case analysis,verifying the feasibility and effectiveness of the proposed method.