语义图引导的非侵入式负荷监测

Semantic Graph Guided Nonintrusive Load Monitoring

摘要: 目的：针对分钟级总表数据中的设备状态混叠与累计能耗偏差，提出语义图引导的非侵入式负荷监测方法。方法：融合低频谱时特征与设备关系先验，构建图谱偏置注意力，联合预测设备状态和分项功率，以总表上界约束抑制超额分解，并通过原型校准复核异常事件。结果：在REFIT、UK-DALE和BLOND上，状态宏平均F1分别达到91.7%、93.1%和87.4%，功率平均绝对误差分别为20.4、15.6和7.8 W；相较于潜在抽象桥Transformer，F1提高1.2、0.5和1.7个百分点。消融实验表明，图谱偏置改善状态区分与功率估计，上界约束降低累计能耗偏差。结论：设备关系与电气观测具有互补性，联合计量约束能够提高低频负荷监测精度，为分项用能分析提供支持。

Abstract: Objective: To address ambiguous appliance states and cumulative energy-consumption deviations in minute-level aggregate power data, this study proposes a semantic graph guided method for nonintrusive load monitoring. Methods: Low frequency spectral-temporal features are integrated with prior knowledge of appliance relationships to construct graph-biased attention for joint prediction of appliance states and disaggregated power, with an aggregate-power upper-bound constraint suppressing over-disaggregation and prototype calibration supporting the review of anomalous events. Results: On REFIT, UK-DALE, and BLOND, the state macro-averaged F1 scores reach 91.7%, 93.1%, and 87.4%, respectively, and the corresponding mean absolute power errors are 20.4, 15.6, and 7.8 W; compared with the latent abstraction bridge Transformer, F1 increases by 1.2, 0.5, and 1.7 percentage points, respectively. Ablation experiments show that graph biases improve state discrimination and power estimation, while the upper-bound constraint reduces cumulative energy-consumption deviations. Conclusion: Appliance relationships and electrical observations are complementary, and incorporating the aggregate-power upper-bound constraint can improve the accuracy of low frequency load monitoring and support appliance-level energy-use analysis.