基于XGBoost算法的电力系统暂态稳定评估

针对暂态稳定评估问题的特点,在改进极限梯度提升(XGBoost)算法的基础上进行暂态稳定评估。根据电网物理特点,定义能够反映电力系统稳态运行状态的特征集;研究XGBoost算法用于暂态稳定评估的过程:针对暂态稳定预测中2类错误严重程度不同的特点,定义包含注意力系数的对数损失函数,使得模型对不稳定样本的误预测情况减少;使用Logistic函数将模型输出概率化,用于衡量XGBoost模型输出的可靠程度,预防部分误预测;给出针对任意系统随机产生样本集的方法。IEEE 39节点系统仿真结果表明,XGBoost算法在准确率上均高于其他几类常用机器学习算法,优化后的损失函数降低了不稳定样本错误分类的可能性,使该算法的召回率较优于其他方法,且概率化输出的形式有助于评估模型输出的可靠程度,降低了误预测的概率。

Transient stability assessment of power system based on XGBoost algorithm

The TSA(Transient Stability Assessment) is carried out based on the improved XGBoost(eXtreme Gradient Boosting) algorithm according to the characteristics of TSA problem. The feature set that can reflect the stable operation state of power system is defined based on the physical characteristics of power grid. The TSA process with XGBoost is researched. Since the severities of two mistakes in TSA process are different, a logarithmic loss function containing the attention coefficient is defined to reduce the mistaken prediction situations of the model to unstable samples. The Logistic function is introduced to transform the model output into a probabilistic form, which is used to measure the reliability of model output to prevent part of mistaken prediction. A method to generate sample sets for any power system is given. The simulative results of IEEE 39-bus power system show that the accuracy of XGBoost is higher than that of other commonly used machine learning methods, the misclassification probability of unstable samples is reduced by the loss function after optimization, which makes the recall rate of the proposed algorithm better than other methods, and the probabilistic output format is helpful for evaluating the reliability of model output and reduces the probability of mistaken prediction.