Intelligent Fault Detection, Classification, and Location in Three-Phase Power Systems using Artificial Neural Networks
- lms editor
- 11 hours ago
- 4 min read
Abstract
The operational integrity of modern power grids is mainly depend on the speed and accuracy of transmission line protection system. In this paper, a robust intelligent framework is proposed for automatic detection, classification and location of faults in a multi-source three-bus power system using Artificial Neural Networks (ANN). The proposed method utilize single-ended measurements taken at Bus 1, where root-mean-square (RMS) values of three-phase voltages and currents are extracted along with zero-sequence components and . These features effectively captures both symmetrical and unsymmetrical transient signatures.
A dual-purpose 5-bit target vector is adopted, where first four bits are used for binary fault classification and the fifth output represent the continuous distance estimation of the fault location. MATLAB/Simulink validation results shows high generalization capability, achieving regression coefficient values up to 0.99. The predicted distance error remain very small, for example a 4 km fault is estimated as 3.8 km. The results proves that ANN-based protection scheme offers scalable and high-precision alternative compared to conventional impedance based relays, thus improving grid resiliency and operational reliability.
Keywords
· Artificial Neural Networks
· Fault Classification
· Fault Location
· Power System Protection
· Transmission Line
2. I. Introduction
In modern smart grid infrastructure, rapid and accurate fault identification become extremely critical for system stability. Due to increasing penetration of distributed and renewable energy sources, transient behavior of the grid become more complex and non-linear. Therefore, conventional protection relays based on fixed impedance setting often fails to generalize properly under multi-source operating conditions.
Traditional distance relays works based on apparent impedance calculation:
where = Phase voltage = Phase current
However, during complex faults and infeed conditions, this apparent impedance may deviate significantly, leading to under-reach or over-reach issues.
Artificial Neural Networks (ANN) provide a data-driven alternative. ANN possess strong non-linear approximation ability and high dimensional pattern recognition capability. By learning fault signatures from voltage and current patterns, ANN can classify and estimate location without explicit impedance calculation.
The objective of this paper is to develop a comprehensive ANN-based framework for:
1. Fault detection
2. Fault classification
3. Fault location estimation
in a three-bus multi-source transmission system using single-ended measurements.
3. II. System Configuration and Methodology
A three-bus transmission system is considered for evaluating the proposed ANN model. Bus 1 and Bus 2 act as source buses, while Bus 3 act as load bus. A 10 km transmission line connects Bus 1 to Bus 3, which is considered as primary protection zone.
3.1 Input Feature Extraction
The ANN input vector consists of 8 features:
Where RMS values are calculated as:
Zero sequence components are given by:
These zero-sequence components helps in identifying ground involved faults (LG, LLG).
Table 1: System Configuration Parameters
Parameter | Specification |
System Topology | Three-bus (Bus 1, 2, 3) |
Source Buses | Bus 1 and Bus 2 |
Load Bus | Bus 3 |
Transmission Line Length | 10 km |
Monitoring Location | Bus 1 (Single-ended) |
Input Features | RMS , RMS , , |
Training Line Lengths | 1 km, 5 km, 10 km, 20 km |
4. III. Neural Network Architecture and Control Strategy
A feed-forward multilayer perceptron (MLP) network is employed. The network structure consist of:
· Input layer: 8 neurons
· Hidden layer: 10–15 neurons (tuned experimentally)
· Output layer: 5 neurons
4.1 Target Encoding
A 5-bit output vector is used:
Where:
· → Binary classification bits
· → Continuous distance estimation (km)
Example encoding:
Condition | Target Vector |
Normal | 0000 0 |
AG Fault | 1001 d |
AB Fault | 1100 d |
ABC Fault | 1111 d |
4.2 Detection and Decision Logic
Detection threshold:
Binary decoding:
Distance estimation:
Training is performed using Levenberg-Marquardt backpropagation algorithm minimizing Mean Square Error (MSE):
Where = Network output = Target value
5. IV. Simulation Model and Parameters
The complete model is developed in MATLAB/Simulink environment. Different fault categories are simulated:
· Line-to-Ground (LG): AG, BG, CG
· Double-Line-to-Ground (LLG): ABG, ACG, BCG
· Line-to-Line (LL): AB, AC, BC
· Three-Phase (LLL): ABC
Multiple line lengths (1 km, 5 km, 10 km, 20 km) are used for ensuring spatial generalization capability.
The fault distance is varied as:
Where = Total line length
6. V. Results and Discussion
The ANN performance is evaluated using regression analysis.
Regression coefficient:
A high indicate strong correlation between predicted and target values.
Table 2: Comparison of Actual vs Predicted Fault Locations
Fault Type | Actual Distance (km) | Predicted Distance (km) | Error (km) |
ABG | 4.0 | 3.80 | -0.20 |
ABG | 6.0 | 5.76 | -0.24 |
ABG | 9.0 | 8.80 | -0.20 |
BG | 9.0 | 9.09 | +0.09 |
Error is calculated as:
It can be observed that maximum error remain within ±0.24 km, which is acceptable for practical field implementation. Slight underestimation observed in ABG cases, while BG case shows small overestimation.
The classification accuracy remain above 98%, showing the network effectively distinguish between symmetrical and asymmetrical faults.
7. VI. Conclusion and Future Scope
This paper presents an intelligent ANN-based protection framework for fault detection, classification and location in three-phase power system. Using only single-ended measurements at Bus 1, the proposed model eliminates the need for synchronized communication infrastructure.
The regression performance reaching up to 0.99 and distance estimation error within ±0.24 km demonstrate that ANN approach provides highly reliable and scalable solution for modern smart grids.
Future work will focus on:
· Extending model for looped and mesh network topologies
· Incorporating deep learning methods such as CNN or LSTM
· Evaluating performance under noisy measurement condition
· Hardware implementation using real-time digital simulator
Overall, the proposed ANN protection scheme significantly enhances system reliability and reduce restoration time in practical transmission networks.
VII. YouTube Video
VIII. Purchase link of the Model
SKU: 0131
Comments