Metering data from the advanced metering infrastructure can be used to find abnormal electricity behavior for the detection of electricity theft, which causes huge financial losses to electric companies every year. This article proposes an electricity theft detector using metering data based on extreme gradient boosting (XGBoost). The metering data are preprocessed, including recover missing and erroneous values and normalization. The classification model based on XGBoost is trained using both benign and malicious samples. Simulations are done by using the Irish Smart Energy Trails data set with six certain attack types. Compared with the support vector machine, decision tree, and other eight machine learning methods, the proposed method can detect electricity theft with either higher accuracy or lower false-positive rate. Experiment results also demonstrate that the proposed method is robust when the data are imbalanced. Our codes are available at https://github.com/WenHe-Hnu/Electric_Theft_XGBoost .

Electricity theft has been a growing concern for the smart grid. It can be defined as follows: illegal customers use energy from electric utilities without a contract or manipulate their meter readings to pay less or not pay the electricity bill. Over the past decade, significant studies have been done to prevent and combat theft. This article aims to provide a general overview of the progress of electricity theft detection, including threat models, datasets and input features adopted, methodologies and techniques, and evaluation metrics. We also make a performance comparison for each detection method. Finally, the challenges and future research directions are presented in this article.

The wide bandwidth of phase-locked loop (PLL) will increase the negative real part of the output impedance of the grid-connected inverter (GCI), thus destroying the stability of the weak grids. This article proposes a novel method to improve the system stability through decoupling the PLL and grid impedance. Initially, the coupling relationship between PLL and grid impedance is revealed. It can explain why the choice of PLL bandwidth and the grid impedance magnitude is limited. Then, the decoupling method is proposed where the primary control structure of GCI does not need to change and uses the structured voltage, the product of PCC voltage and grid admittance, as the improved input of PLL. The proposed method can achieve the full-frequency passive of the output impedance of GCI even if the PLL has high bandwidth and the impedance measurement error is up to 60%, which means the adverse stability effect of PLL on the system is eliminated. The theoretical analysis is verified by simulations and experiments.

Practical applications of non-intrusive load monitoring (NILM) require load recognition models that generalize to unseen data from new houses and operate efficiently on edge devices. However, existing NILM approaches, particularly deep learning models, are computationally intensive and prone to performance degradation when deployed to new houses due to domain shifts. To address these challenges, this article proposes a Weighted Transferable Random Forest (WTRF) approach for load recognition. Based on the Random Forest (RF) framework, WTRF incorporates a transfer learning mechanism to swiftly adapt the model to new houses using only 1 to 3 labeled samples per appliance. The model is lightweight, with a memory size under 300 KB. Case studies on three datasets demonstrate its effectiveness, including a macro F1-score of 97.0 ± 2.6% when transferring from PLAID to WHITED, a significant improvement over 5.7±1.7% achieved by source-only models. Deployed on a Raspberry Pi 4, WTRF achieves update times as low as 3.1±0.3 seconds and testing times of approximately 3 ms per house. These results highlight WTRF’s efficiency in addressing domain shifts and its suitability for real-time NILM in resource-constrained edge environments.

Non-Intrusive Load Monitoring (NILM) identifies individual appliance power usage within an overall power load, enabling more refined and secure load management. However, existing deep learning-based NILM models require large amounts of labeled data from diverse devices, which is time-consuming and raises privacy concerns. Additionally, handling these large datasets demands significant computational and memory resources. To address these issues, we propose a self-alignment, source-aware domain adaptation approach. Our method employs domain adversarial networks to address feature and label distribution shifts between source and target domains. To preserve privacy, we fine-tune the model without source domain data. To stabilize adversarial training, we incorporate a self-alignment mechanism (SAM). The SAM ensures parameter updates without accessing source domain data, enabling stable training while preserving privacy. Confidence-based label density maps generate pseudo-labels for fine-tuning. We validated our approach with intra-domain and inter-domain adaptability studies on synthetic and real data. We conducted intra-domain and inter-domain adaptability studies on synthetic and real data. Results show our method achieves decomposition accuracy superior to source-based methods for devices with regular usage patterns, all while effectively preserving privacy by eliminating the need for source data during the fine-tuning stage. This offers potential for improving NILM efficiency and energy management in industrial measurement settings with similar stability requirements.

Recently, transfer learning has gained attention in non-intrusive load monitoring (NILM) as it improves the generalization of the models, particularly when transferring knowledge from one real-world dataset to another. Considering the privacy concerns arising from model transfer training due to the large-scale collection of annotated real-world data, synthetic datasets have become a viable alternative for improving generalization performance in NILM. However, the generalization in research scenarios with domain distribution gap, like synthetic-to-real settings, still needs exploration. In this paper, we present a feature reconstruction-based network model for NILM, designed to capture both common shared and unique representations across synthetic and real data domains. An external attention mechanism is adopted to compensate for the lack of measurement data, utilizing synthetic datasets to address both the scarcity of annotated data and privacy concerns. Ultimately, the analysis of inter-domain data features influences the model’s generalizability. Our experimental results demonstrate that the proposed network shows promise for the transfer learning from a synthetic dataset to a real dataset, and its performance relies on the power consumption of appliance and the dissimilarity of probability mass functions between the two datasets.