对于新型电力系统,提前预测系统惯量水平是消除系统惯量薄弱风险的重要前提。而预测所使用的机器学习模型大多是“黑箱”模型,存在可解释性不足的问题。为此,提出了一种基于可解释性极端梯度提升算法 (extreme gradient boosting,XGBoost)的电力系统惯量短期预测方法。该方法根据系统惯量响应特性分析,选择电力系统运行数据和气象数据作为输入特征。基于SHAP归因方法构建可解释性XGBoost的解释机制,通过计算Shapley值来量化各特征的重要程度,从而对模型预测结果进行多维度解构分析。以实际系统为例进行实证分析,结果表明,所提方法能够有效地预测电力系统短期惯量以及解释特征对预测的影响。

Predicting the level of in advance in new power systems is essential to eliminate the risk of a weak system inertia, and black-box machine learning models, which have insufficient interpretability, are widely used for system-inertia predictions. Therefore, this paper introduces a short-term prediction method, based on interpretable extreme gradient boosting (XGBoost), for power system inertia. Based on the analysis of the system inertia response characteristics, the method selects the power system operation and meteorological data as input features. The interpretation mechanism of XGBoost was constructed based on Shapley additive explanation values. By calculating the Shapley value to quantify the importance of each feature, the model prediction results can be deconstructed into multiple dimensions. Simulations were performed using a realistic photovoltaic system, and the results showed that the proposed method can effectively predict the short-term inertia of a power system as well as elucidate the influence of the features on the predicted results.

分析了大规模新能源电力系统按分区配置惯量的必要性,并对此给出电压源型换流器采用功率同步控制以配置虚拟惯量的建议。据此,提出一种考虑频率变化率约束的惯量分区配置方法。该方法先将待研究系统按频率响应特性的一致程度进行分区,然后结合分区内部扰动、分区间联络线上功率变化和负荷电压响应等因素计算扰动情况下各分区的最大不平衡功率,从而以分区为单位得到惯量需求并进行配置。最后,基于华东电网规划算例,说明了苏北沿海电网等分区的惯量配置需求,并验证所提惯量分区配置方法的有效性。

The necessity of configuring inertia regionally for large-scale renewable energy power system is analyzed. To achieve this goal, a suggestion is put forward, which says that the power synchronization control can be utilized to configure virtual inertia for voltage source converters. Based on these, a regional inertia configuration method considering a constraint on the rate of change of frequency （RoCoF） is proposed. In this method, firstly, the system to be studied is partitioned according to the consistency of frequency response.Then, the maximum unbalanced power of each region under disturbance is calculated by combining the internal disturbance in the region, the change of the power on transmission lines between different regions and the load response to the disturbed voltage. After that, the inertia demand could be obtained and the inertia configuration could be carried out regionally. Finally, in the case study of East China Power Grid, the inertia requirement of each region in the system is illustrated, and the effectiveness of the proposed regional inertia configuration method is verified.

大规模风电并网降低了电力系统惯量水平,增加了其暂态频率偏差的越限风险,然而现有频率预测模型对含风电系统的在线预测速度和精度都还不够,故需进一步优化频率偏差极值预测方法,用以系统的频率稳定评估。文章基于广域量测技术(wide area measurement system, WAMS),考虑风电并网对频率响应过程的影响,提出了一种物理-数据融合频率偏差极值在线预测方法。首先,利用广域量测信息对有功-频率开环解耦模型进行数据粘合,在此基础上形成可快速求解的物理-数据融合的暂态频率分析模型;其次,基于该模型,获得实时更新的频率偏差极值预测值,并提出“预测值误差指数”指标来量化预测精度,指导在线模型的自适应动态结果输出;最后,通过算例验证了所提频率偏差极值在线预测方法的快速性和准确性。

A large-scale wind power grid connection reduces the level of inertia in a power system and increases the out-of-limit risk of transient frequency deviations. However, the existing frequency prediction models are not suitable for a fast and highly accurate online prediction of wind power system characteristics. Therefore, further optimizing the extreme value prediction method that predicts frequency deviations is necessary for evaluating the frequency stability of such systems. This paper introduces an online prediction method for predicting the frequency deviation extremum based on physical data fusion; this method is based on a wide-area measurement system and considers the influence of the wind power grid connection on the frequency response process. First, a transient frequency analysis model of physical data fusion was developed based on data bonding of the open-loop decoupling model of active power and frequency using wide-area measurement information. Second, the predicted extreme value of frequency deviation is updated in real-time using the model. A “prediction value error index” is proposed to quantify the prediction accuracy and guide the adaptive dynamic output of the online model. Finally, an example is provided to verify the prediction speed and accuracy of the extremum frequency deviation predicted by the proposed online prediction method.