隐私保护下基于XGBoost-LightGBM融合模型用电数据异常检测

Privacy-Preserving Electricity Consumption Anomaly Detection Based on XGBoost-LightGBM Hybrid Model

摘要: 据Northeast Group报导，全球由于窃电造成的经济损失高达960亿美元/年，对发达国家和发展中国家用电安全均产生了严重影响，为了减少窃电损失，亟需提升用户侧用电数据异常检测的安全性和效率。本研究在传统的电力数据异常检测框架上引入基于AES + RSA的混合加密架构，并结合SHA-256加密哈希算法和数字签名技术实现数据传输的安全性保护、完整性验证与身份认证；在异常检测阶段，利用网格搜索对XGBoost与LightGBM模型进行参数优化后，通过构建以AUC与预测差异性为动态权重调整因子的模型融合式，使优化后的XGBoost与LightGBM模型实现自适应融合以提升检测方法的泛化性。利用国家电网公开数据集进行检测实验，结果显示该模型AUC达到了82.03%，Accuracy为91.53%，G-mean值为54.99%，继而在4个KEEL公开数据集上进行泛化性能测试，结果表明该检测方法具有较好的检测异常样本的能力。

Abstract: According to a Northeast Group report, electricity theft causes up to USD 96 billion in annual global economic losses, severely compromising power security in both developed and developing nations. To mitigate these losses, enhancing the security and efficiency of user-side power-consumption anomaly detection is imperative. This study extends the conventional anomaly-detection framework by integrating a hybrid encryption scheme based on AES and RSA, combined with SHA-256 hashing and digital-signature techniques, to ensure data confidentiality, integrity verification, and authentication during transmission. In the anomaly-detection phase, the research first employs grid search to optimize hyperparameters for XGBoost and LightGBM models. The authors then fuse the optimized models via a dynamic weighting mechanism where weights are adaptively adjusted based on each model’s AUC and prediction diversity, thereby improving the ensemble’s generalization capability. Experimental evaluation on a publicly available State Grid dataset demonstrates that the proposed hybrid model achieves an AUC of 82.03%, an accuracy of 91.53%, and a G-mean of 54.99%. Generalization tests on four KEEL benchmark datasets confirm the method’s robust anomaly detection capability.