针对光伏发电功率随机性强、波动性大导致其预测精度不高的问题,提出一种基于自适应近邻传播聚类(adaptive affinity propagation clustering,adAP)、多模式分解、多分支输入组合的光伏功率预测方法.首先,基于相关性分析找到与光伏发电功率高度相关的气象因素,并利用快速傅里叶变换(fast Fourier transform,FFT)将光伏输出功率从时域转换到频域,与相关度高的气象因素一起作为 adAP 算法的聚类特征,对具有相似气象特征的日场景进行分类;其次,对聚类相似日较少且输出功率波动剧烈天气类型中的气象相关因素和光伏输出功率添加高斯白噪声,并将其与原始数据合并,达到倍增样本的效果,以提升模型的泛化能力和鲁棒性;然后,使用变分模态分解(variational mode decomposition,VMD)、奇异谱分解(singular spectrum decomposition,SSD)和群分解(swarm decomposition,SWD)对光伏功率、辐照度和温度进行分解,削弱原始序列的波动性,丰富模型的输入特征;最后,搭建多分支的残差网络(residual network,ResNet)和长短期记忆网络(long short term memory network,LSTM)模型,提取数据的时间特征和波动特征,合并后输入到门控循环单元网络(gated recurrent unit network,GRU)中,建立历史特征和未来光伏输出功率的联系,得到预测结果.实验结果表明,所提出的多模型组合预测方法在光伏功率波动较缓天气情况下,能够保持较高的预测精度;在波动剧烈天气情况下,能够较大地提升预测精度.
Ultra-short-term Prediction of Photovoltaic Power Based on Multi-mode Decomposition and Multi-branch Input
To address the problem that photovoltaic(PV)power is highly stochastic and volatile leading to its low pre-diction accuracy.We propose a PV power prediction method based on adaptive affinity propagation clustering(adAP),multi-mode decomposition,and multi-branch input combination.First,the meteorological factors which are highly corre-lated with PV power generation are found based on correlation analysis,the PV power generation is converted from the time domain to the frequency domain by using the fast Fourier transform(FFT),which is used as the clustering feature of the adAP algorithm along with the meteorological factors with high correlation,and the daily scenes with similar meteor-ological characteristics are classified.Secondly,Gaussian white noise is added to the meteorological correlation factors and PV output power in the clustered weather types with fewer similar days and intense output power fluctuations,which is merged with the original data to achieve the effect of multiplying the samples and to improve the model's generalization ability and robustness.Then,the variational mode decomposition(VMD),singular spectrum decomposition(SSD)and swarm decomposition(SWD)are used to decompose PV power,irradiance and temperature to weaken the volatility of the original series and enrich the input features of the model.Finally,a multi-branch residual network(ResNet)and a long short term memory network(LSTM)model are built to extract the temporal and fluctuating features of the data,which are combined and input into a gated recurrent unit network(GRU)to establish the connection between the historical features and the future PV output power,and get the prediction results.The experimental results show that the proposed mul-ti-model combination prediction method can maintain high prediction accuracy in the case of slow fluctuation of PV power,and improve prediction accuracy in the case of severe fluctuation of weather.
photovoltaic power generationultra-short-term forecastadaptive affinity propagation clusteringmulti-branch inputmulti-mode decompositiondeep learning
万方数据
维普
