Sep 2026· IEEE Transactions on Smart Grid· Vol 17, pp. 4802-4815· 0 citations· 47 references
Abstract
This paper proposes a fault location method for DC distribution networks (DCDNs) based on graph neural networks (GNNs), which integrates the fault line selection (FLS) and fault distance estimation (FDE) that are conventionally handled independently. The proposed method focuses on the analysis of feeders, including FLS of multiple feeders and FDE of a single feeder. Specifically, a feeder-as-node graph is constructed, where synchronous measurement data are extracted as node features, ensuring consistent dimensionality and enhanced learning efficiency. Moreover, the proposed method explicitly embeds the logical interdependencies between FLS and FDE into the structural design and parameter updating mechanism. An output processing module is designed to estimate the fault distance by analyzing the FLS results, ensuring the model utilizes the data of the DCDN system-level information rather than a single feeder. Furthermore, a two-stage pre-training strategy is introduced to improve stability and generalization, in which partial parameters are frozen. The effectiveness and generalization of the proposed method are verified by hardware-in-the-loop experiments.
Accurate and robust fault location is of great importance for improving the power supply reliability and automation level of modern distribution networks. However, existing methods often rely on line parameters or signal features and fail to fully exploit physical topology priors and dynamic fault-response relationships. To address this issue, this paper proposes a Physical-Information Dual-Graph Adaptive Fusion Network (PIDGAF-Net) for distribution network fault location. The proposed method constructs a static physical graph to encode feeder topology and electrical coupling, and a dynamic information graph to characterize sample-specific node response similarity under fault conditions. A dual-branch graph attention encoder and a node-level gating mechanism are then employed to adaptively fuse the two graph representations, enabling more discriminative fault feature learning. Experimental results on the IEEE 33-bus distribution system show that PIDGAF-Net achieves the highest localization accuracy of 95.87%, while maintaining moderate computation time, parameter scale, and memory overhead. In addition, the proposed method remains stable under inaccurate measurement, asynchronous sampling, data loss, and mixed anomaly conditions. These results demonstrate its strong potential for intelligent fault diagnosis and operation support in distribution networks.
Zhengkai Sun, Qian Zhang· Journal of Physics, Conferen...· 0 citations
: With the carbon peaking and carbon neutrality targets, the penetration of distributed renewable energy such as photovoltaic (PV) systems in distribution networks is increasing. However, the rapid growth of rooftop PV causes operational problems such as transformer overloading and voltage violations. These issues challenge the distributed PV integration capability of distribution networks. At the same time, distribution networks have more complex load types, more flexible topologies, and frequently changing parameters. It is still difficult to achieve accurate and efficient assessment. To address this problem, this paper proposes a message-passing graph neural network (MPNN)-based method for assessing distributed photovoltaic integration capability in distribution networks. The proposed framework leverages historical operational data from multiple operating states. It applies the message-passing graph neural network to learn node topology features. It also uses a gated recurrent unit to capture temporal transitions across different topologies. An improved optimization algorithm is used to improve efficiency. The method considers practical constraints under complex and changing network conditions. It identifies bottlenecks of renewable integration and helps optimize resource allocation. It can improve the distributed PV integration capability of distribution networks. Case studies show that the method can achieve fast and accurate assessment of the maximum integration capability of distributed PV systems, which can provide support for the safe and economic operation of distribution networks.
Ning Guo, Jian Liu, Haixiang Zang et al.· Energy Engineering· 0 citations
Aiming at the problems of dynamic topology and strong output uncertainty in distribution networks caused by highpenetration new energy integration, as well as the core pain points of traditional situation deduction methods including low computational efficiency, incomplete scenario coverage, and insufficient deduction accuracy under extreme scenarios, this paper proposes a fast operation situation deduction method for high-penetration new energy distribution networks integrating conventional and extreme scenarios based on graph computing.Firstly, a physical-information fusion attribute graph model of the distribution network is constructed to realize the unified representation of topology, electrical characteristics and uncertainty information, as well as multi-scale dynamic topology adaptation. Secondly, a conventional-extreme scenario fusion modeling framework is established. Through scenario normalization, K-means clustering reduction and graph attention network (GAT) correlation learning, key representative scenarios covering the core characteristics of the two types of scenarios are screened out.Then, a storage-computation integrated graph computing core engine is designed based on the power supply service panorama graph. Combined with spatio-temporal parallel deduction and incremental reuse mechanism of time-evolution graphs, the deduction efficiency is greatly improved.Meanwhile, a GCN-LSTM fusion model is introduced to realize multi-step situation prediction and accurate risk identification of distribution networks.Experiments are carried out on a county-level distribution network with 1200 nodes and 52% new energy penetration. Compared with the traditional Newton-Raphson method and GCN-based deduction method, the results show that the single-scenario power flow deduction time of the proposed method is only 800 ms, and the parallel deduction time of 25 key scenarios is 2.3 s. The efficiency is improved by more than 50% compared with the traditional method, and the state estimation speed is increased by 83.3% due to the incremental reuse mechanism.The voltage/power flow prediction error of the proposed method is lower than that of the comparison methods, with a scenario coverage of 95% and a risk identification accuracy of over 96%. The method can be seamlessly integrated with the distribution network automation system, and effectively supports the conventional operation monitoring and extreme scenario emergency disposal of high-penetration new energy distribution networks, with significant theoretical value and engineering application prospects.
Gang Wang, He Wang, Lin Peng et al.· International Conference on...· 0 citations
Accurate fault location is critical for distribution network reliability. However, increasing distributed energy resource (DER) penetration complicates fault location due to intermittent generation and bidirectional power flows that reshape fault signatures. Spatio-Temporal Graph Neural Networks (STGNNs) have shown promise by jointly modeling spatial and temporal dependencies, but their behavior under increasing DER penetration has not been studied rigorously. In this paper, we (i) systematically benchmark spatio-temporal graph attention network (STGATv2) against purely temporal (gated recurrent unit, GRU), purely spatial (GATv2) and traditional machine learning baselines, and (ii) evaluate how well models generalize across increasing DER penetration levels (10%, 25%, 50%) on a reconfigured IEEE 123-bus feeder with multiple DER injection points and moderate-to-high impedance faults. Results show that STGATv2 consistently outperforms neural baselines, achieving 92-94% macro F1 in-distribution. Notably, generalization across penetration levels is asymmetric: training at 50% penetration retains near in-distribution F1 score at lower levels, whereas training at 10% degrades considerably at 50% - with STGATv2 retaining 81-84% F1 under these drastic shifts, substantially higher than GATv2 and GRU which drop to 69-74% F1 and 73-75% F1 respectively. Under realistic measurement noise, STGATv2 maintains>85% F1, while GRU drops as low as 33.5% F1, highlighting the critical role of topological awareness for robust fault location in active distribution networks.
Burak Karabulut, Olayiwola Arowolo, Carlo Manna et al.· 0 citations
The modern agricultural equipment needs to be able to cope with changes in field conditions, however, fault detection is still challenging for the time being with the traditional rule based or single sensor based diagnostic procedures. These methods frequently do not reflect data interactions and interdependencies within components. To overcome these drawbacks, this research introduces Multi-Source Graph-Enhanced Fault Localization (MS-GEFL) framework. The framework combines vibration sensors, operational logs, environmental data and controller outputs in a multi-source fusion layer to guarantee that the data is consistent. Then, a dynamic graph model is used to represent the mechanical system and Graph Neural Networks (GNNs) are used to propagate the information related to the fault across different operating conditions to identify the anomalies. The experimental results show that the precision, recall and F1 of MS-GEFL are 93.9%, 94.6% and 94.0% respectively. Additionally, the framework makes operations more efficient, with an average localization delay of 1.28 seconds and an average robustness index of 91.7% even in the presence of noise in sensor data. This technology is intelligent and reliable solution for fault localization in complex agricultural machinery.
Gaofeng Wu, Han Zhang, Yunquan Li· Diagnostyka· 0 citations
This study develops a low-complexity data-driven method for real-time fault classification, and localization on a 20 kV radial feeder, while accounting for the computational and memory constraints of embedded hardware. The proposed approach is motivated by the changes introduced by distributed generation (DG), which alters fault-current behavior and reduces the reliability of conventional impedance-based protection. These limitations are mainly caused by reverse power flows and the additional infeed contribution from DG units during fault conditions. The proposed hybrid architecture processes time- and frequency-domain features, including symmetrical components and wavelet energy, through a dual-stage inference chain. The first stage uses a Random Forest classifier for fault-type identification, while the second stage uses a Multi-Layer Perceptron (MLP) distance estimator for fault localization. Evaluation on 23,778 MATLAB/Simulink fault scenarios, was performed using a grouped scenario-level holdout split to reduce overlap between training and test cases. The proposed approach reduces the localization error compared with the conventional impedance method, which produced errors exceeding 14 km under DG operation. The Random Forest classifier achieved 99.44% fault-type accuracy, and the MLP estimator reached a Mean Absolute Error (MAE) of 93.8 m, with localization accuracy of 99.51% within 1 km. Analytical resource profiling estimates less than 300 KB Flash and under 50 µs execution time, assuming a Cortex-M4F/M7-class target, pending hardware-in-the-loop validation.
M. Lemkharbech, S. Sarih, Z. Boulghasoul et al.· International Conference on...· 0 citations
Related blog posts
MIT News · Artificial Intelligence· news.mit.eduJul 15, 2026
Assistant Professor Pat Pataranutaporn describes a new interface that lets everyday users glimpse inside an AI's neural network before their chatbot ever says a word.
MIT News · Artificial Intelligence· news.mit.eduJul 6, 2026
PhD student Rachel Sava, winner of the Envisioning the Future of Computing Prize, explores transformative improvements and dystopian risks of neural technology.