基于包络线聚类的多模融合超短期光伏功率预测算法

针对传统功率预测方法以气象因素进行聚类划分时各气象因素权重难以分配以及单模型预测精度较差的问题,提出一种基于光伏功率包络线聚类的多模融合超短期光伏功率预测算法.对异常特征数据进行预处理,采用Pearson相关系数与XGB Feature Importance模块分析光伏功率和各特征之间的相关关系,并构建新特征;介绍包络线理论,并根据光伏功率包络线参数进行聚类划分,将聚类后的数据作为输入,借鉴Stacking集成学习框架构造XGBoost+LightGBM+LSTM融合模型对光伏功率进行预测;将所提算法与气象因素聚类和功率区间聚类下的各预测算法进行实验对比;为了避免训练过程中模型超参数的影响,采用K折交叉验证对数据的训练集、验证集和测试集进行划分.仿真结果表明,所提算法较气象因素和功率区间聚类法能有效提高复杂天气情况下光伏功率预测精度,且多模融合效果总体优于单独算法模型.

Multi-mode fusion ultra-short-term photovoltaic power prediction algorithm based on envelope clustering

Aiming at the problems that the weights of meteorological factors are difficult to assign and the prediction accuracy of single model is poor when traditional power prediction methods are used for meteorological factor cluster division, a multi-mode fusion ultra-short-term photovoltaic power prediction algorithm based on envelope clustering of photovoltaic clustering is proposed. The abnormal feature data is preprocessed, Pearson correlation coefficient and XGB Feature Importance module are adopted to analyze the correlation between photovoltaic power and each feature, and new features are constructed. The envelope theory is introduced, the photovoltaic envelope parameters are used for cluster division, the data after clustering is taken as input, and the fusion model of XGBoost + LightGBM + LSTM is constructed using Stacking integrated learning framework to predict photovoltaic power. The prediction algorithm is compared with each forecasting algorithm under meteorological factor clustering and power interval clustering. In order to avoid the influence of model hyperparameters during training process, K fold cross validation is used to divide the training set, validation set and test set of the data. The simulative results show that the proposed algorithm can effectively improve the prediction accuracy of photovoltaic power under complex weather conditions compared with the meteorological factor and power interval clustering methods, and multi-mode fusion effect is totally better than single algorithm model.