考虑多气象因子累积影响的光伏发电功率预测

Photovoltaic Power Generation Forecasting Considering the Cumulative Effects of Multiple Meteorological Factors

针对当前主流的光伏发电功率预测方法中,深度学习算法训练耗时长、传统机器学习算法精度有待提升的问题,提出考虑多气象因子累积影响的光伏发电功率预测方法。首先,采用相关系数法筛选关联气象因子,并使用余弦距离的K-means++算法将训练集高效划分为K个类簇;在第1路预测中,使用关联气象因子构造二维气象矩阵,然后输入到柯西变异优化的特征金字塔网络(feature pyramid networks,FPN)模型,挖掘相关气象因子对光伏发电功率的累积影响;在第2路预测中,使用轻量梯度提升机(light gradient boosting machine,LightGBM)算法实现光伏发电功率的即时预测;借鉴集成学习的思想,将上述2组预测结果加权求和,得到最终的光伏发电功率预测结果。在关联因子筛选、聚类效果对比实验部分,取0.3为相关系数阈值,验证聚类个数取4为最优;在光伏出力预测算法对比实验部分,分别基于平均相对误差、均方根误差,计算所提算法的预测精度为88.12%、82.03%,均高于其他各项参照算法,从而证明了所提算法的可行性。

Aiming at the problems that training time of deep learning algorithm is long,and the accuracy of traditional machine learning algorithm needs to be improved in current mainstream photovoltaic generated power prediction methods,a photovoltaic generated power prediction method considering cumulative impact of multiple meteorological factors is proposed.Firstly,this paper uses the correlation coefficient method to screen associated meteorological factors,and K-means++algorithm with cosine distance to efficiently divide training set into K clusters.In the first path of prediction,the paper constructs a two-dimensional meteorological matrix by using related meteorological factors,and then inputs into Cauchy mutation optimized feature pyramid network(FPN)to mine cumulative impact of related meteorological factors on photovoltaic generated power generation.In the second path of prediction,the light gradient boosting machine(LightGBM)algorithm is used to realize immediate prediction of photovoltaic generated power.Borrowing the idea of ensemble learning,the above two prediction results are weighted and summed to obtain final photovoltaic power prediction results.In the experimental parts,the paper takes 0.3 as threshold of correlation coefficient and verifies that 4 is the optimal number of clusters.Based on the average relative error and root mean square error,the prediction accuracy of this algorithm is 88.12%and 82.03%,which are higher than other reference algorithms,proves the feasibility of proposed method.