MATLAB Simulation of Power System Fault Detection and Classification Using Deep Neural Network

Introduction to Power System Fault Detection and Classification

Fault detection and classification are essential to safeguard power system equipment from damage caused by faults such as line-to-ground, line-to-line, and three-phase faults. Accurate and timely identification allows protective devices to isolate faulty sections and prevent widespread outages.

Using machine learning techniques, especially deep neural networks, helps automate this process with high accuracy by analyzing electrical signals collected from the power system.

Overview of the Power System Model

The simulation model consists of an 11 kV, 30 MVA power source connected to a load via a step-down transformer that reduces the voltage from 11 kV to 400 V for consumer usage. Faults are introduced at the 11 kV bus to simulate various fault conditions.

To gather data for the neural network, the model measures the RMS voltage and current for each phase, along with zero-sequence voltage and current components. These signals serve as inputs for the DNN, while target values are assigned depending on whether the system is operating normally or under fault conditions.

Data Collection for Training the Neural Network

Data must be collected under both normal and fault conditions for effective training. Multiple fault types are modeled, including:

• Single line-to-ground faults (e.g., A-G, B-G, C-G)

• Line-to-line faults

• Double line-to-ground faults

• Three-phase faults

Simulations consider different fault resistance values (0.1 Ω, 10 Ω, 100 Ω) to account for various fault severities. Each scenario is run for 4 seconds, and the measured RMS and zero-sequence values are recorded as input features for the neural network.

Fault Detection Model

The fault detection network is trained to classify the presence or absence of a fault. The output target is set to zero during normal operation and one when any fault occurs. This binary classification allows the model to quickly determine whether a fault exists.

After gathering all input-output data pairs, the data is used to train the neural network using MATLAB’s Neural Network Fitting Tool. The dataset is split into training, validation, and testing subsets to ensure the model’s robustness.

The network training yields a very low mean squared error and an R-value close to 1, indicating an excellent fit between input data and fault detection output.

Integration into Simulink for Real-Time Detection

Once trained, the neural network model is saved and integrated into a Simulink environment. This allows the system to process real-time input data from the power system and output fault detection status continuously.

The Simulink model receives eight input signals (RMS voltages and currents, zero-sequence components) and produces a single output indicating fault presence or absence.

Fault Classification Model

Beyond detection, classifying the type of fault is critical for appropriate remedial action. A separate neural network is trained for this purpose, using a multi-output binary encoding scheme to represent different fault types.

For example:

• A-G fault might be encoded as 1001

• B-G fault as 0101

• C-G fault as 0011

• Other fault combinations have their unique binary representations

This classification network uses the same input signals but outputs a 4-bit target vector, each bit representing a fault category.

Training and Testing the Fault Classification Network

The classification network is trained similarly, with the dataset split into training, validation, and testing parts. The training achieves high accuracy, with regression coefficients around 97–98%, showing that the model reliably maps input signals to the correct fault type.

The trained model is also integrated into Simulink, where it receives inputs and outputs a classification result in real time.

Simulation Results and Performance Evaluation

The integrated Simulink system is tested under different fault scenarios sequentially:

• Normal operation correctly registers no faults.

• Single line-to-ground faults (AG, BG, CG) are detected and classified correctly.

• Complex faults like line-to-line and double line-to-ground faults are also identified with high accuracy.

• Out of 11 simulated fault types, 10 were classified correctly.

• The model misclassified the ABC ground fault as an ABC fault, slightly lowering overall accuracy.

The fault detection accuracy is near 100%, while fault classification accuracy stands at approximately 91%.

Conclusion

This MATLAB/Simulink implementation of power system fault detection and classification using a deep neural network demonstrates an effective approach to protecting power systems. With high detection accuracy and strong classification performance, such models can be integrated into modern protective relaying systems to enhance reliability and safety.