基于滑动分块百分位数Bootstrap法的风电功率概率区间预测

PROBABILISTIC INTERVALS FORECASTING OF WIND POWER BASED ON MOVING BLOCK PERCENTILE BOOTSTRAP METHOD

提出一种基于非参数滑动分块百分位数Bootstrap法(MBPB)的风电功率概率区间预测方法。由于风功率数据存在显著的时间相依结构,该方法首先对预测功率进行等间隔划分,再以某区间内的预测误差序列为样本,借助滑动分块Bootstrap法(MBB)抽样产生多个伪样本,然后对伪样本数据通过滑动分块百分位Bootstrap法和四分位法相结合的统计推断生成一定置信水平下的误差上下限,进而得到该预测功率段内的概率预测区间。同时建立包含区间覆盖率和区间平均带宽的评价指标,通过将其与百分位法、百分位数Bootstrap(PB)法的预测结果对比,表明基于MBPB的概率性预测区间的覆盖率更高,平均带宽更窄,精度更好且效果也更优。

A probabilistic wind generation forecasting intervals was established based on non-parametric Moving Block Percentile Bootstrap(MBPB)method. Due to the temporal dependence in power data,firstly,prediction power is divided equidistantly. Then moving block bootstrap method is applied to reshape new samples using forecasting errors of each power section. Combining MBPB and Quartile together,the lower upper bounds estimation under given confidence level can be obtained by utilizing the combination method for deduction. Finally,probabilistic forecasting interval of corresponding power section is achieved,and evidently superior to both Percentile and Percentile Bootstrap(PB)methods in terms of Prediction Interval Coverage Probability and PI Normalized Average Width which are introduced to assess the quality of interval. Results show that the proposed approach ensures a higher PICP and a narrower PINAW simultaneously,indicating a more satisfactory performance.