The distribution company Elektro Ljubljana d.d., operates and maintains 17,470 km medium and low-voltage electricity distribution network supplying 356,106 end users. While the medium-voltage network is remotely monitored and controlled through the SCADA Advanced Distribution Management System (ADMS), the low-voltage network remains significantly less observable due to its size, dispersion, and large number of transformer stations, overhead lines, cables, and metering points. 

As electrification and network complexity continue to increase, rapidly identifying faults and abnormal operating conditions in the low-voltage network has become increasingly challenging. With the goal of enabling predictive maintenance and improve network reliability, Elektro Ljubljana d.d. developed a machine learning–based fault detection approach using 10-minute averaged voltage measurements and events from smart electricity meters. 

Challenge

Detecting faults in highly branched low-voltage distribution networks is challenging due to the delayed availability of measurement data, typically received with a one-day delay, large numbers of distributed assets, and limited real-time observability. 

Many low-voltage anomalies — including missing phases, broken neutral conductors, poor cable contacts, blown fuses, and voltage irregularities — are difficult to identify using conventional monitoring approaches and are often detected only after customer complaints or on-site inspections. 

In addition, the large volume of smart-meter measurements and recorded events makes manual analysis impractical for maintenance teams operating across multiple distribution units. 

Elektro Ljubljana d.d. therefore required a scalable and data-driven solution capable of automatically identifying abnormal network behavior and supporting predictive maintenance activities using real operational data from the low-voltage distribution network. 

Solution

Elektro Ljubljana d.d. developed a supervised machine learning solution in MATLAB for automated fault detection in low-voltage distribution networks. 

The algorithm analyses weekly smart-meter measurements, including three-phase voltages and recorded meter events. Real anomalies collected by maintenance teams — such as broken neutral conductors, blown fuses, missing phases, and poor cable contacts — were used to design and validate the detection methodology. 

The machine learning workflow combines:

• automated data filtering,
• statistical feature extraction,
• feature selection and PCA dimensionality reduction,
• supervised anomaly classification,
• iterative feedback from maintenance teams.

Figure 1: Workflow of the machine learning–based fault detection solution for low-voltage distribution networks. 

On the training dataset (90 % of the data), 3-fold cross-validation was applied to mitigate overfitting and avoid capturing noise and random fluctuations. Multiple machine learning algorithms were evaluated, including weighted kNN (k-nearest neighbours), SVM (support vector machines), linear discriminant, Naïve Bayes, decision trees and neural network. Weighted kNN with Euclidean distance metric and optimized hyperparameters achieved the most robust performance across different testing scenarios.

The developed algorithm continuously scans weekly measurements using a 1-hour sliding window and classifies anomalies together with confidence levels and criticality indicators before forwarding detected events to maintenance teams.

Results

The machine learning solution is currently deployed as an automated weekly analysis tool for the entire low-voltage distribution network of Elektro Ljubljana d.d., processing smart-meter measurements and events from more than 356,000 end users.

The algorithm successfully detects multiple real-world anomalies, including:

• bad contacts on neutral conductors,
• missing phases,
• blown fuses and
• neutral conductor failures.

Several detected anomalies were confirmed and resolved by on-site maintenance teams, including faults that were difficult to identify using conventional monitoring approaches.

The system currently detects an average of 170 anomalies per week across the entire low-voltage network. More critical faults, such as missing neutral conductors or poor neutral-line contacts, are typically detected at a rate of up to 2 faults per distribution transformer station.

Maintenance teams use the algorithm as a complementary predictive-maintenance tool for early fault identification and prioritization of field inspections.

The long-term objective is deployment of the developed solution on a production server with automated reporting and integration into operational maintenance workflows.

Learn More

About The Author