How I Used Artificial Neural Networks to Detect Transmission Line Faults

For my final year project, I built and tested an ANN-based fault detection and classification system (FDANN) for underground transmission lines, comparing its performance directly against the classical impedance-based method that most protection schemes still rely on today. This is a walkthrough of how the system was structured and what the results actually showed.

The Starting Question

Impedance-based fault location is simple, well-understood, and cheap to implement — which is exactly why it's still the default in most protection relays. But it has a known weak point: accuracy degrades sharply as fault resistance increases, because the underlying assumption (that a fault behaves close to a dead short) breaks down. High-resistance faults — from degraded insulation, partial breakdown, or high-impedance ground contact — are common in aging underground networks, which is exactly where the classical method struggles most. The project set out to test whether a trained neural network could hold accuracy where the impedance method falls apart.

Building the Dataset

Since real-world labeled fault data at the volume needed for training is hard to come by, the dataset was built from simulation — modeling the underground cable feeder with varying fault types (line-to-ground, line-to-line, double line-to-ground, three-phase), fault locations distributed along the line length, and fault resistances ranging from near-zero (bolted fault) up to 100 Ω to specifically stress-test the high-resistance region where classical methods weaken.

Each simulated fault event produced a three-phase current and voltage waveform capture, from which features were extracted rather than feeding raw waveforms directly into the network:

• RMS values of the three phase currents and voltages during the fault
• Harmonic content ratios, since fault arcing and non-linear fault paths introduce characteristic harmonic signatures
• Fault inception angle relative to the voltage waveform
• Sequence component magnitudes (positive, negative, zero sequence) — these are particularly informative for classifying fault type

This feature extraction step matters more than it might seem — raw waveform samples make for a much larger and noisier input space, and physically meaningful features (especially sequence components, which protection engineers already use conceptually) give the network a much more learnable representation of what actually distinguishes fault types and locations.

Network Architecture and Training

The system used two separate feedforward ANN stages rather than one combined network: a classification stage to first identify the fault type, followed by a regression stage to estimate fault distance — splitting a mixed classification-and-regression problem into two better-defined sub-problems rather than forcing one network to do both simultaneously. Both stages used backpropagation training with a standard train/validation/test split, with the validation set used to prevent overfitting given the moderate dataset size from simulation.

Results

Across the test set, the fault-type classification stage achieved 97.8% accuracy — correctly identifying which phases and ground were involved in the fault across all four fault type categories.

The distance estimation stage was benchmarked directly against the classical impedance-based method across the same test scenarios, specifically at the high-resistance end (Rf = 100 Ω) where the impedance method is known to be weakest:

• FDANN (ANN-based): 3.6% mean absolute error in fault distance estimation
• Classical impedance method: 24.3% mean absolute error under the same conditions

That gap is the core finding of the project. At low fault resistance, both methods perform reasonably, because the dead-short assumption underlying the impedance method roughly holds. As fault resistance rises, the impedance method's error grows steeply while the trained network's error stays comparatively flat — because it isn't relying on that single physical assumption in the first place, just on patterns learned from the training data.

Limitations Worth Being Honest About

The results come with real caveats. The training and test data were simulation-based rather than drawn from field fault records, which means real-world performance depends on how well the simulation model reflects actual network behaviour — cable ageing characteristics, joint impedances, and load variability that a simulation can approximate but not fully replicate. A production deployment would need retraining (or at least validation) against real fault records specific to the network it's protecting, and the model's performance on fault scenarios outside its training distribution — an unusual combination of location and resistance not well represented in training — is inherently less predictable than a physics-based method's behaviour on the same edge case.

Why This Matters Beyond the Project

This wasn't meant to be a claim that ANN methods should replace impedance relays outright — that's neither realistic nor necessary. The more useful framing is that ANN-based estimation is well suited as a complementary tool layered alongside existing protection: impedance relays continue handling fast primary protection decisions, while an ANN-based system provides a more accurate secondary estimate specifically for locating high-resistance faults during the restoration process, where speed matters less than accuracy and every percentage point of distance error translates directly into metres of unnecessary excavation.