A streaming computer system identified and classified simulated grid faults with up to 96.4% accuracy.
The system processes incoming data from phasor measurement units, smart meters and monitoring and control systems. It performs some initial processing near the devices that collect the data, then uses distributed stream-processing software and machine-learning models to detect unusual behavior and classify possible faults.
In a simulated 33-bus IEEE test network, the system predicted injected faults in under a second, with accuracy reaching 96.4%. The researchers also report that it remained scalable and fault tolerant as data loads and the placement of edge-computing resources changed.
What the system found
The proposed system combined near-device data processing, distributed streaming tools and two machine-learning approaches: long short-term memory networks and random-forest classifiers. The models were updated as new data arrived, allowing the system to detect anomalies and classify faults with low delay.
When tested on a simulated 33-bus IEEE power network with injected faults, it achieved sub-second prediction accuracy of up to 96.4%. The abstract also reports that the system remained scalable and fault tolerant under changing data loads and edge-deployment conditions.
Why faster fault warnings matter
Fault diagnosis that works on live data could give grid operators information sooner than systems based on offline, batch analysis. The researchers say this could support proactive maintenance and help energy-management systems respond to problems before they disrupt operations.
The result is relevant to increasingly networked power infrastructure, but the study shows this performance in a simulated test network rather than in a live power grid.
// Source
Progress in Electronics and Communication Engineering · 2026 · DOI: 10.31838/ece/03.02.03
Authors: Ishrat Zahan Mukti, Ebadur Rahman Khan
Institutions: Independent University, Bangladesh
