基于CNN特征选择与QRGRU的电力负荷概率密度预测方法

Power Load Probability Density Prediction Method Based on CNN Feature Selection and QRGRU

针对一般预测方法难以提供负荷概率性信息,且难以兼顾负荷数据的时序性和天气、日类型等非连续特征的缺陷,文章提出一种基于卷积神经网络(Convolutional Neural Network,CNN)结合门控循环神经网络分位数回归(Gated Recurrent Neural Network Quantile Regression,QRGRU)的概率密度预测方法。将历史负荷值、天气及日类型等信息利用滑动时间窗构造连续特征图,通过CNN提取不同类型特征的潜在信息,以此作为QRGRU输入,预测不同分位点下未来一天任意时刻负荷的预测结果,并通过核密度估计获得任意时刻负荷的概率密度分布。实验结果表明,采用CNN学习不同类型特征与负荷需求的关系,并结合QRGRU和核密度估计技术,可以更好地解决电力负荷概率密度预测问题。

Traditional forecasting methods are difficult to provide load probability information,and it is difficult to take into account the time sequence of load data and the discontinuous characteristics of weather and day type.A probability density prediction method based on convolutional neural network(CNN)combined with gated recurrent neural network quantile regression(QRGRU)is proposed.In this paper,historical load values,weather,and day types are used to construct continuous feature maps using sliding time windows.The potential information of different types of features is extracted through CNN as QRGRU input to predict load at any time in the future under different quantile.The probability density distribution of the load at any time can be obtained by kernel density estimation.Experimental results show that the use of CNN to learn the relationship between different types of features and load demand,combined with QRGRU and kernel density estimation technology,can better solve the problem of power load probability density prediction.