TL;DR
This paper introduces a novel Transformer Neural Network (PT-TNNet) model, enhanced with data fusion techniques, for improved fault diagnosis of power transformers within intelligent substation one-key sequential control systems. The PT-TNNet demonstrates superior accuracy in monitoring transformer health, addressing critical issues like data redundancy and slow response times in existing systems.
What Problem Does This Paper Address?
Intelligent substations are increasingly complex, generating vast amounts of data. This surge in data, coupled with the limitations of traditional manual operations, necessitates advanced automated systems. The "one-key sequential control system" aims to streamline operations and enhance safety, but it faces challenges: low data transmission efficiency, prolonged system response times, poor data quality and accuracy, and insufficient equipment fault monitoring and diagnosis. These issues can lead to operational errors, increased costs, and reduced grid reliability. The paper specifically targets the critical task of power transformer fault diagnosis within these systems, aiming to improve their overall performance and reliability.
Methods
The core of this research is the proposed Power Transformer-Transformer Neural Network (PT-TNNet) model. This model leverages the power of Transformer neural networks, known for their ability to capture long-range dependencies and complex relationships within data through self-attention mechanisms.
Key methodological components include:
- Data Fusion: Before feeding data into the neural network, a data fusion module is employed. This process combines time-domain and frequency-domain features (obtained via Fourier transform) to reduce data redundancy, enhance data quality, and accelerate model convergence.
- Transformer Architecture: The PT-TNNet utilizes a multi-head self-attention mechanism and a feed-forward network, enabling it to process sequential data effectively and learn intricate patterns indicative of transformer health.
- Feature Extraction and Fusion: The model incorporates distinct modules for feature extraction and fusion, mapping different data sources into a common feature space before classification.
- Comparative Analysis: The PT-TNNet's performance is rigorously compared against various machine learning algorithms, including Random Forest and Extra Trees, to validate its superiority.
- Cross-Validation: To ensure robustness and generalization, cross-validation is performed on key datasets.
Datasets
The study utilizes a dataset collected from sensors within and around power transformers. This dataset comprises measurements of:
- Transformer winding temperature
- Insulation oil temperature
- Ambient temperature
- Three-phase voltage and current
The data was collected over a period from September 2023 to February 2024, with readings taken every 15 minutes. The paper provides a breakdown of samples during normal operation, fault occurrence, and invalid samples. No new benchmark datasets were introduced; existing data from transformer monitoring was used.
Key Insights and Ideas
- Transformer Networks Excel at Temporal Dependencies: The self-attention mechanism in Transformer models is particularly well-suited for capturing the complex temporal and inter-dimensional relationships inherent in power transformer operational data, outperforming traditional models in this domain.
- Data Fusion is Crucial for Efficiency: Combining time-domain and frequency-domain features through data fusion significantly improves model performance by expanding feature dimensions, reducing noise, and accelerating convergence, which is vital for real-time applications.
- Internal Transformer Health is Key: While ambient temperature is a factor, the study highlights that winding temperature and insulation oil temperature are the most critical indicators for accurate transformer condition monitoring and fault diagnosis.
- PT-TNNet Offers a Robust Solution: The proposed PT-TNNet model demonstrates a significant leap in accuracy and reliability for detecting and diagnosing transformer faults, directly addressing the shortcomings of current one-key sequential control systems.
Key Results
The PT-TNNet model achieved impressive results across various metrics:
- Winding Temperature: Accuracy of 95.68%, Precision of 94.25%, Recall of 95.46%, and F1 Score of 94.85%.
- Oil Temperature: Accuracy of 99.95%, Precision of 99.89%, Recall of 95.45%, and F1 Score of 97.61%.
- Ambient Temperature: Accuracy of 97.96%, Precision of 96.72%, Recall of 94.65%, and F1 Score of 95.67%.
In most cases, the PT-TNNet outperformed other tested algorithms, including Random Forest and Extra Trees, particularly for winding and oil temperature data. Cross-validation further confirmed the model's robustness and generalization capabilities. The paper concludes that the PT-TNNet is feasible for application in one-key sequential control systems, enhancing their performance.
Limitations and Future Work
- Model Complexity vs. Response Time: The paper acknowledges that as the depth of the PT-TNNet model increases for improved accuracy, its complexity also rises. This could be a drawback for systems requiring extremely rapid responses.
- Ambient Temperature Variability: The performance on ambient temperature data was not as optimal as for internal parameters, attributed to its susceptibility to complex and fluctuating natural variations.
- Data Availability: The study notes that the data supporting the findings is unavailable due to privacy concerns.
- Future Work: While not explicitly detailed, the authors suggest that further increasing the depth of the PT-TNNet model could continue to improve accuracy, implying a trade-off with computational resources and response time that warrants further investigation. The focus on thermal faults could also be extended to other types of transformer faults.
Summary updated