基于LightGBM算法的光伏并网系统孤岛检测及其集成的可解释研究

针对智能孤岛检测方法欠缺对数据集划分过程中标签分布不均问题的考虑，以及该领域尚未对复杂智能孤岛检测模型的决策进行可解释性分析，提出了一种基于轻梯度提升机(LightGBM)算法的孤岛检测模型。采用分层抽样的K折交叉验证检测模型的分类性能，解决数据标签分布不均的问题；提出基于决策树的Shapley值加性解释方法为主干，融合累计局部效应图和局部代理模型的集成可解释分析框架，从全局性和局部性角度对光伏并网系统的孤岛状态检测进行归因分析。算例仿真结果表明，所提模型能在传统检测方法的检测盲区中实现精确且快速的动态孤岛检测，且在电压波动、系统故障等情况下均未发生误判。基于集成的归因分析方法解决了单一可解释方法的欠合理性问题，揭示了模型输入电气特征自变量与孤岛检测响应因变量之间的关系，提高了模型的可信度。

Islanding detection of grid-connected photovoltaic system based on LightGBM algorithm and its integrated interpretability analysis

In view of the lack of considering the uneven label distribution in the data set partitioning process of intelligent islanding detection method and the lack of interpretability analysis for the decision of complex intelligent islanding detection models, an islanding detection model based on light gradient boosting machine(LightGBM) algorithm is proposed. The classification performance of the model is tested by K-fold cross-validation with stratified sampling to solve the uneven distribution problem of data labels. Taking the decision tree based Shapley additive explanation(TreeSHAP) method as the main work, the integrated interpretability analysis framework combining accumulated local effect graphs and local agent models is proposed to analyze the attribution of islanding state detection for the grid-connected photovoltaic system from the global and local perspectives. The simulative results show that the proposed model can achieve accurate and fast dynamic islanding detection in the detection blind area of traditional detection methods, and no misjudgment occurs in the cases of voltage fluctuation and system fault. The integration-based attribution analysis method solves the unreasonableness problem of the single interpretability method, reveals the relationship between the input electrical characteristic independent variables and the response dependent variables of islanding detection, and improves the reliability of the model.