稀疏高斯过程在短期风电功率概率预测中的应用

Application of sparse gaussian process in short-term wind power probability prediction

针对短期风电功率预测,提出一类基于稀疏高斯过程(sparse gaussian processes,Sparse-GP)的概率预测方法。通过对数据集随机划分所形成的数据子集,给出基于数据点子集(subset of datapoints,SoD)近似、回归子集(subset of regressors,SoR)近似、投影过程(projected process,PP)近似算法的3种Sparse-GP方法,该方法不仅能给出模型的均值预测,而且能获取模型的预测方差,这很好地解释了模型置信水平。不同的Sparse-GP方法在保持常规GP方法优点的同时,还能解决GP方法随着训练数据增加而产生的矩阵运算困难等难题,且计算效率高。将具有不同协方差函数形式的Sparse-GP方法应用于不同地区的短期风电功率单步与多步预测实例中,在同等条件下还与常规GP、SVM方法进行对比。实验结果表明,Sparse-GP方法可以给出较好的预测效果,且适用于较大规模数据集的训练。

For short-term wind power prediction,a class of a class of probabilistic prediction methods based on sparse gaussian processes (Sparse-GP) is proposed. Three different Sparse-GP methods using subset of datapoints(SoD) approximation,subset of regressors(SoR) approximation and projected process(PP) approximation algorithm were respectively given by randomly dividing the data set into a subset of data. The Sparse-GP methods can give the mean prediction of the model,and obtain the prediction variance of the model,which explains the confidence level of the model well. Different Sparse-GP methods can solve the difficulty of matrix operation caused by the increase of training data in the GP method while maintaining the advantages of the conventional GP method,and the calculation efficiency is high. The employed Sparse-GP methods with different covariance functions were then applied to the instances for single-step and multi-step prediction of short-term wind power in different region,and compared with the conventional GP and SVM methods under the same conditions. Experimental results show that the Sparse-GP methods can give better prediction results and are suitable for training large-scale data sets.