Intelligent Transformer Fault Diagnosis: A Deep Learning Application of Dissolved Gas Analysis

Preface

Power transformers are critical and costly assets in the electric power infrastructure. Their failure can lead to significant disruptions and costly repairs. Early fault detection is essential to maintain reliable power system operation. Online monitoring systems offer a solution by enabling real-time data collection on transformer health parameters. By proactively identifying and addressing issues, utilities can significantly improve transformer reliability and lifespan, ultimately reducing maintenance costs and minimizing service disruptions. By making use of online monitoring and advanced diagnostic techniques, utilities can make informed decisions to optimize transformer performance and minimize the risk of catastrophic failures.

Methodology

Dissolved Gas Analysis (DGA) is a powerful diagnostic tool for identifying incipient faults in oil-filled power transformers. By measuring the concentration of specific gases hydrogen, methane, ethane, ethylene, and acetylene. DGA can pinpoint the type of fault, whether it's electrical or thermal.

When a transformer experiences electrical stress, such as partial discharge or arcing, it can lead to the breakdown of insulation and the generation of gases like hydrogen, ethylene, and acetylene. Conversely, thermal stress, resulting from overheating, can cause the decomposition of cellulose insulation, producing carbon monoxide and carbon dioxide.

To interpret DGA results, various methods have been developed, including the Roger's Ratio method, the Doernenburg Ratio Technique, and the Duval Triangle method etc.. Among these, the Roger's Ratio method, standardized in IEEE C57.104-2019, is widely used. This method calculates specific gas ratios to identify fault types, providing valuable insights into the health of the transformer. By employing DGA and advanced interpretation techniques, utilities can proactively monitor transformer condition, schedule preventive maintenance, and avoid interruption of power supply.

Predicting Transformer Life with Deep Learning

Predictive maintenance of power transformers is revolutionized by the integration of Dissolved Gas Analysis (DGA) with deep learning. By analyzing historical DGA data, neural networks can accurately predict the remaining lifespan of transformers and identify potential issues early on. This proactive approach significantly reduces the risk of unexpected failures, minimizes downtime, and optimizes maintenance schedules. Deep learning models, typically comprising multiple layers, process input features such as gas ratios, age, and maintenance intervals. Through a series of feedforward operations, activation functions, and backpropagation, the model learns complex patterns in the data and generates precise predictions. By leveraging the power of deep learning, utilities can ensure the reliable and efficient operation of their power systems.

To prepare the data for model training, we first one-hot encode the categorical fault type column to create numerical representations for each fault type. Then, we separate the features X &Y. To ensure that features are on a similar scale, we apply standardization using Standard Scaler. Finally, we split the data into training and testing sets, reserving 20% of the data for testing and using the remaining 80% for training the model.

from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.model_selection import train_test_split

df = pd.get_dummies(df, columns=["Fault_Type"])
X = df.drop(columns=["Fault_Type_PD", "Fault_Type_DLE", "Fault_Type_DHE", "Fault_Type_Thermal1", "Fault_Type_Thermal2", "Fault_Type_Thermal3"])
y = df[["Fault_Type_PD", "Fault_Type_DLE", "Fault_Type_DHE", "Fault_Type_Thermal1", "Fault_Type_Thermal2", "Fault_Type_Thermal3"]]

scaler = StandardScaler()
X_scaled = scaler.fit_transform(X)

X_train, X_test, y_train, y_test = train_test_split(X_scaled, y, test_size=0.2, random_state=42)        

Imports the TensorFlow library, providing tools for building and training machine learning models. Imports specific components from Keras, including layers for building neural network architectures and models for organizing and training these architectures and define the model.

import tensorflow as tf
from tensorflow.keras import layers, models
model = models.Sequential([
    layers.Dense(64, activation='relu', input_shape=(X_train.shape[1],)),
    layers.Dense(32, activation='relu'),
    layers.Dense(16, activation='relu'),
    layers.Dense(y_train.shape[1], activation='softmax')  # Softmax for multiclass classification
])
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
history = model.fit(X_train, y_train, epochs=50, batch_size=8, validation_data=(X_test, y_test))        

To gain further insight into the training process, we visualize the model's learning curve using Matplotlib. The script plots both the training accuracy and validation accuracy over epochs.

import matplotlib.pyplot as plt

# Evaluate model
test_loss, test_accuracy = model.evaluate(X_test, y_test)
print(f"Test Accuracy: {test_accuracy:.2f}")

# Plot training history
plt.plot(history.history['accuracy'], label='accuracy')
plt.plot(history.history['val_accuracy'], label = 'val_accuracy')
plt.xlabel('Epoch')
plt.ylabel('Accuracy')
plt.legend(loc='lower right')
plt.show()        

This allows us to observe how the model's performance changes throughout training and identify potential issues like overfitting.

Conclusion

This deep learning model for transformer fault classification based on DGA data demonstrates promising results in accurately identifying various fault types. By taking advantage of the power of neural networks, the model effectively extracts intricate patterns and relationships of DGA data, enabling precise fault diagnosis. The model's ability to handle complex data and learn from large datasets makes it a valuable tool for proactive maintenance and condition monitoring of power transformers. Further R&D can be explored, techniques like transfer learning to enhance the more accuracy in the performance of the model. Also, incorporating real-time data streams and integrating the model into advanced monitoring systems can revolutionize transformer maintenance practices.

Reference:

1. 3D Implementation of Rogers Method for Power Transformer Oil Condition Monitoring Metin VARAN1 , Lamine Cherif N. MOUCHILI 2 1Asst. Prof. Dr., Faculty of Technology-Department of Electrical and Electronics Engineering, 2 Faculty of Computer and Information Sciences, Computer Engineering Department Sakarya University, Esentepe Campus 54100-Sakarya/Turkey.
2. IEEE Guide for the Interpretation of Gases Generated in Oil-Immersed Transformers; IEEE: Piscataway, NJ, USA, 2008
3. Basuki, A.; Suwarno. Online Dissolved Gas Analysis of Power Transformers Based on Decision Tree Model. In Proceedings of the Conference on Power Engineering and Renewable Energy (ICPERE), Solo, Indonesia, 29–31 October 2018

#utilityanalytics #digitaltransformation #businessintelligence #forecasting #Disruptiveinnovation #datascience #machinelearning #energyanalytics #artificialintelligence #dataanalytics #deeplearning #tensorflow #EDA #predictivemaintenance #statisticalanalysis #powerdistribution #utility #python #advancedanalytics #transformer #DGA #keras #electricalengineering #powersystem