基于XGBoost的双层协同实时校正超短期光伏预测

针对超短期光伏预测应对突发过程性天气时准确性普遍下降,而通过实时气象监测校正辐射值对设备要求较高、对精细化预报依赖性强的问题,以数据驱动为理念,提出了基于XGBoost的双层协同实时校正超短期光伏预测模型。根据大气运动的连续演变性和自相似性,从机器学习角度推演气象整体连续变化的过程,提升预测精度。首先,基于数值天气预报(NWP),在基准层建立强相关气象特征的预测模型。然后,在实时层由临近时段内的基准层动态预测情况挖掘潜在的气象变化规律,并推测未来预测时段气象因素对于光伏出力的影响,对时段内基准预测值进行逐点校正。采用中国杭州滨江一实际光伏电站实采数据进行算例分析,分别与基于NWP特征学习、时序分析、误差推移的XGBoost预测模型以及决策树、支持向量机、长短期记忆网络这3种经典预测模型相比较,结果表明所提模型具有更高的超短期光伏预测精度。

XGBoost Based Bi-layer Collaborative Real-time Calibration for Ultra-short-term Photovoltaic Prediction

The accuracy of ultra-short-term photovoltaic (PV) prediction generally declines when facing sudden processing weather, and radiation value calibration through real-time meteorological monitoring has high equipment requirements and a strong dependence on refined prediction. To solve the problems above, this paper proposes a bi-layer collaborative prediction model with real-time calibration based on extreme gradient boosting (XGBoost) by the guidelines of data-driven concept. According to the continuous evolution and self-similarity of atmospheric motion, the model deduces the overall meteorological change process from the perspective of machine learning to improve the prediction accuracy. Firstly, based on numerical weather prediction (NWP), a prediction model is established on the basic layer based on highly correlated meteorological features. Secondly, on the real-time layer, the dynamic prediction of the basic layer in the adjacent time period is used to explore the potential meteorological changes, and the impact of meteorological factors on the PV output in the future prediction time period is speculated. The reference prediction values in this time period are corrected point by point. Through the case study of the measured data in a certain PV station in Binjiang District, Hangzhou, China, the results show that the proposed model has higher accuracy for ultra-short-term PV power predicition, compared with the XGBoost prediction model based on NWP feature learning, time series, error shifting, and three classical prediction model like decision tree, support vector machine (SVM), long short-term memory (LSTM) network.