基于降噪时序深度学习网络的风电功率短期预测方法

Short-term Wind Power Forecasting Method Based on Noise-reduction Time-series Deep Learning Network

利用风电场历史功率数据预测未来一段时间内的风功率,对保障电网安全稳定运行具有重要的意义。本文提出一种基于奇异谱分析SSA(singular spectrum analysis)和长短时记忆LSTM(long-short term memory net⁃work)网络的时序特征预测框架用于短期风功率的预测。首先通过SSA对历史风功率原始数据进行降噪处理,然后经过数据转换之后,以LSTM网络为基础进行预测模型的训练,最后通过某风电场提供的两个风机的历史功率数据进行验证。实验结果表明,奇异谱分析对风电场的历史数据具有良好的降噪性,SSA+LSTM模型在测试数据上取得了较好的预测性能,能够有效进行短期风功率的预测。

Using the historical power data from a wind farm to predict wind power in the future is of significance for en⁃suring the safe and stable operation of power grid.In this paper,a time-series feature forecasting framework is proposed for the short-term wind power prediction,which is based on singular spectrum analysis(SSA)and long-short term memory(LSTM)network.First,the raw data of historical wind power are denoised by SSA.Then,the forecasting mod⁃el is trained based on the LSTM network after the data conversion.Finally,the forecasting method is verified by the his⁃torical power data of two fans provided by one wind farm.Experimental results show that SSA has a satisfying noise-re⁃duction performance for the historical power data from the wind farm.Moreover,the SSA-LSTM model has a better pre⁃diction performance on the test data,which can effectively predict the short-term wind power.