基于相似日和分位数回归森林的光伏发电功率概率密度预测

Forecasting of photovoltaic power generation probability density based on similar day and quantile regression forests

为提高光伏发电功率预测精度及可靠性,提出一种基于相似日和分位数回归森林(QRF)的光伏发电功率概率密度预测模型。选取某光伏电站实测数据为研究对象,在将光伏发电功率原始数据按不同天气类型进行分类的基础上,通过温度和风速2个特征向量选取相似日,并对相似日历史数据建立BP神经网络(BPNN)、支持向量机(SVM)和QRF预测模型。结果表明:晴天时,不同模型预测值均能较好跟踪真实值变化趋势,在13:00-16:00光伏发电功率下降时间段,QRF模型更接近真实值;多云或阴天时,在9:00-12:00,3种模型预测误差均较大;雨天时,在14:00-16:00光伏发电功率突变时间段,BPNN模型预测误差最大,SVM预测值相对于QRF模型更接近真实值,而在10:00-12:00,SVM模型预测误差增大。对不同模型不同天气类型下的预测误差,QRF模型预测性能更佳。

In order to improve the forecasting accuracy and reliability of photovoltaic power generation, a photovoltaic power generation probability density forecasting model based on similar day and quantile regression forests(QRF) is proposed. The measured data of a photovoltaic power station are selected as the research object. On the basis of classifying the original data of the photovoltaic power generation according to different weather types, similar days are selected by two eigenvectors(temperature and wind speed). Moreover, the BP neural network(BPNN) model, support vector machine(SVM) model and QRF forecasting model are established for the historical data of similar days. The results show that, on sunny days, the forecasting results of different models can better track the change trend of true values, and the QRF model’s result is closer to the true value during the period of photovoltaic power reduction from 13:00 to 16:00. On cloudy days, the forecasting errors of the above three models are relatively large from 9:00 to 12:00. In rainy days, during the sudden change of photovoltaic power generation from 14:00 to 16:00, the BPNN model has the largest forecasting error, and the SVM model’s forecasting result is closer to the true value than the QRF model, but the forecasting error of the SVM model increases from 10:00 to 12:00. The QRF model has better prediction performance for different weather types among different models.