Graph Neural Network for Distribution Fault Location
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fault location methods in power distribution systems, such as impedance-based and machine learning approaches, fail to accurately detect and locate short circuit faults due to factors like fault type, unbalanced loads, and system topology changes, leading to inefficiencies in power restoration and increased outage durations.
Innovation Solution
A graph neural network (GNN) based fault location method that considers both node and branch attributes, including phase voltages, currents, equivalent conductance, and susceptance matrices, to model system topology and accurately determine fault locations on branches, enabling precise fault detection and classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional machine learning approaches (ANN, SVM, CNN) are used for fault location, then the fault location can be estimated based on bus measurements, but the accuracy deteriorates when system topology or branch parameters change because these approaches ignore topology configurations and branch regulations
Solution Approach 1:
The patent applies dynamics by transitioning from static topology assumptions to dynamic topology adaptation. The GNN model is designed to dynamically adapt to different system topologies and branch configurations through its graph structure, which can represent and process varying network configurations. The model learns to handle topology changes by processing graph representations that encode the current system state, enabling accurate fault location despite dynamic changes in network structure.
Solution Approach 2:
The patent applies parameter changes by incorporating branch parameters (impedance, admittance, regulation settings, energization status) as explicit inputs to the GNN model. These parameters are integrated into the graph representation and used during fault location estimation. When branch parameters change, the model receives updated parameter values and adjusts its fault location calculation accordingly, maintaining accuracy under varying operating conditions.
2Measurement precision
If impedance-based methods are used for fault location, then the fault location can be estimated using voltage and current measurements, but the accuracy deteriorates due to fault type variations, unbalanced loads, heterogeneity of overhead lines, and measurement errors
Solution Approach 1:
The patent applies the intermediary principle by introducing a GNN-based intermediate processing layer between the raw measurements and the fault location estimate. The GNN model serves as an intelligent intermediary that processes voltage, current, and branch parameter measurements, automatically extracting relevant features and compensating for measurement errors, unbalanced loads, and fault type variations. This neural network intermediary learns to handle various harmful factors during training and provides robust fault location estimates under diverse operating conditions.
Solution Approach 2:
The patent applies composite materials by combining multiple types of measurements and parameters into a unified GNN model input. The model integrates voltage measurements, current measurements, branch impedance parameters, admittance parameters, regulation settings, and energization status into a composite representation. This composite approach allows the model to leverage complementary information from different sources, improving accuracy while compensating for individual measurement limitations and reducing the impact of various harmful factors.
3Measurement precision
If traveling wave-based methods are used for fault location, then the fault location can be determined by observing original and reflected waves, but the complexity increases due to requirements for high sampling rates, communication overhead, or additional signal injection
Solution Approach 1:
The patent applies mechanics substitution by replacing the physical signal injection and high-rate sampling requirements of traveling wave methods with a computational approach. Instead of requiring additional hardware for signal injection and high-speed data acquisition, the GNN model processes standard voltage and current measurements using learned patterns. This substitutes complex mechanical and hardware requirements with intelligent algorithms, achieving comparable or superior accuracy while significantly reducing device complexity and operational requirements.
4Measurement precision
If conventional learned relationship approaches are used, then the fault location can be estimated based on given configuration data, but the computation effort increases tremendously when data adjustments or topology changes require re-learning
Solution Approach 1:
The patent applies preliminary action by pre-training the GNN model on a comprehensive dataset that encompasses diverse system topologies, branch configurations, and operating conditions. The model learns generalizable fault location patterns during an initial training phase that covers a wide range of scenarios. This preliminary training enables the model to handle topology changes and data adjustments without requiring extensive re-learning, as it has already acquired robust fault location capabilities that transfer across different configurations.
Solution Approach 2:
The patent applies universality by designing a GNN model with multi-functional capability to handle various system configurations and fault types. The graph-based architecture and parameter integration enable the model to process different topology structures, branch parameter combinations, and fault scenarios through a unified framework. This universal approach allows the model to maintain accurate fault location performance across diverse conditions without requiring separate models or extensive re-training for each specific configuration.
Data Source
AI summary
Systems and methods performed by a fault detection apparatus for fault detection and localization in distribution feeders having branches and nodes. The method including receive feeder raw data in a feeder of a power system. Process the feeder raw data with given operational electrical characteristics of the feeder to generate a branch attribute dataset for each branch separated by a pair of nodes for all branches. Generate a node attribute dataset for each node for all the nodes in the feeder. Input the branch and node attribute datasets into a trained neural network to determine whether a branch has a fault and a fault location within the branch, to output a classification of the fault and the fault location. Generate an alert signal based upon determining the classified fault and fault location in response to the alert signal to an outage response system.