基于改进高斯过程回归的短期负荷概率区间预测方法

电力市场改革和分布式能源的并网给电网的运行和规划带来了许多不确定性的因素。为获取更准确、更综合的电力负荷预测值信息,提出一种基于K-means特征提取和改进高斯过程回归的短期负荷概率区间预测方法。首先利用历史负荷数据建立候选特征集,然后通过K-means的特征提取方法先对候选特征集进行分类,再利用K邻域内特征变量之间的互信息来选取负荷最优特征子集,并实时更新最优特征子集。为了准确捕捉电力负荷的时变特性,利用改进的高斯过程回归算法进行电力负荷概率区间预测,主要包括动态更新超参数和滑动窗更新训练样本集两个部分。实例表明,所提方法相比分位回归、高斯过程回归而言预测精度更好,所形成的预测区间具有更窄的区间宽度和更高的覆盖率,能为电力系统的运行规划提供更全面、更有效的负荷信息。

Electric load probabilistic interval prediction method based on improved Gaussian process regression

Due to the power market innovation and the reform of integration of distributed energy, many uncertainties factors have been brought to the planning and operation of power grid. In order to obtain more accurate and comprehensive load forecasting information, this paper highlights a probabilistic electric load forecasting method which combines K-means based Feature Selection (KFS) and Improved Gaussian Process Regression (IGPR). Firstly, the candidate feature set is established by using the historical load data. Then the K-means feature extraction method is used to classify the candidate feature set, and the mutual information among the feature variables in the K neighborhood is used to select the load optimal feature subset, as well as update the optimal feature subset along with forecasting process. In order to accurately capture the time-varying characteristics of power load, the improved Gaussian process regression algorithm is used to predict the power load probability interval, which mainly includes two parts:dynamically updating hyperparameter and sliding window to update training sample set. Experiment results show that the proposed model outperforms the Quantile Regression (QR) and Gaussian Process Regression (GPR), which has narrower interval width with higher coverage, thus providing more comprehensive and effective load information for the power system. This work is supported by National Natural Science Foundation of China (No. 51707041) and Science and Technology Project of China Southern Power Grid Company (No. GDKJXM20162087).