基于AdaBoost集成学习的窃电检测研究

针对传统窃电检测中单一分类方法的不足，提出一种基于AdaBoost集成学习的窃电检测算法。首先利用训练集对决策树、误差逆传播神经网络、支持向量机和k最近邻四种方法进行训练对比，提出决策树作为AdaBoost集成学习算法的弱学习器。其次通过绘制不同学习率下的分类错误率曲线，确定AdaBoost集成学习算法的学习率和弱学习器个数。最后利用爱尔兰智能电表数据集中的居民用电数据对所提算法进行测试评估，将AdaBoost集成学习算法与决策树、k最近邻、误差逆传播神经网络、支持向量机等各类单一强学习算法对比。结果表明基于AdaBoost集成学习的窃电检测算法在准确率、命中率、误检率等检测指标中最优，灵敏性分析验证了基于AdaBoost集成学习的窃电检测方法的有效性。

Research on electricity theft detection based on AdaBoost ensemble learning

There is a deficiency in the single classification method in traditional electricity thief detection. Thus a method based on AdaBoost ensemble learning is proposed. First, the training set is used to compare the decision tree, error backpropagation network, support vector machine and k-nearest neighbors, and the decision tree is adopted as the weak learner of the AdaBoost algorithm. Secondly, the learning rate and the number of weak learners of AdaBoost ensemble learning are determined by plotting the error rate curves under different learning rates. Finally, the proposed method is tested and evaluated on the Irish smart meter dataset. It is compared with the single strong learning algorithms, such as decision tree, error backpropagation network, support vector machine, k-nearest neighbors. The results show that electricity theft detection based on AdaBoost ensemble learning is the best among the indicators of accuracy, true positive rate and false positive rate. The sensitivity analysis shows the validity of the electricity theft detection method based on AdaBoost ensemble learning. This work is supported by National Natural Science Foundation of China (No. 51607104) and 2019 Science and Technology Project of State Grid Hubei Electric Power Company (No. 5215K018006B).