TL;DR
This paper introduces a novel method for accurately predicting transformer hot-spot temperatures (HST) even when the three-phase power supply is unbalanced. By combining advanced fluid dynamics simulations with a machine learning approach (Support Vector Regression) and a genetic algorithm for feature selection, the method achieves high accuracy and robustness, crucial for safe and efficient transformer operation.
What Problem Does This Paper Address?
Ensuring the safe and reliable operation of power transformers is paramount. A critical factor is monitoring the hot-spot temperature (HST) of the windings, as excessive heat can lead to accelerated aging and failure. Traditional methods for measuring HST are often invasive or impractical for real-time monitoring. While empirical formulas exist, they can be inaccurate, especially under non-ideal operating conditions. A significant challenge arises when transformers operate under three-phase unbalanced loads, a common scenario in distribution networks. Under these conditions, the location of the hot spot is not fixed to a single phase but shifts dynamically. This dynamic behavior makes it difficult to predict HST using conventional methods, necessitating a more sophisticated approach that can account for these changing thermal dynamics.
Methods
The researchers propose a multi-faceted approach:
- Thermal Fluid Field Simulation: They first employed the finite volume method to create a coupled simulation of the transformer's thermal fluid field. This involves solving complex equations for mass conservation, momentum, and energy to model the intricate flow of cooling oil and heat distribution within the transformer. This provides a detailed understanding of the internal thermal behavior.
- Support Vector Regression (SVR) for Inversion: To predict the HST, an SVR model was utilized. SVR is a powerful machine learning technique known for its ability to model complex, non-linear relationships. The model was trained using data derived from the thermal fluid simulations.
- Orthogonal Experimental Design and Grid Search: To construct robust training and testing datasets, an orthogonal experimental method was used. This ensures that the training data covers a wide range of operating conditions systematically. The SVR model's parameters (penalty coefficient C and kernel parameter γ) were then optimized using a grid search method combined with K-fold cross-validation, a standard technique for hyperparameter tuning.
- Streamline Analysis for Feature Selection: A key innovation is the use of oil flow streamlines to identify representative paths of heat transfer. Under unbalanced conditions, the hot spot shifts, and the oil flow patterns change. By analyzing these streamlines, the researchers identified critical paths and selected corresponding temperature measurement points on the transformer's steel tank as input features for the SVR model.
- Genetic Algorithm for Dimensionality Reduction: To address the challenge of having too many input features (24 steel tank temperature points), a genetic algorithm was employed. This algorithm iteratively optimizes the selection of input features, reducing the dimensionality of the problem while preserving the most important information for HST prediction. This significantly improves the efficiency and operability of the inversion model.
Datasets
The study utilized a simulated dataset generated from the transformer's thermal fluid field numerical calculation for an S13-M-100 kVA/10 kV transformer. This simulation-based dataset was then augmented with experimental data from a temperature rise test for validation. The paper also generated training and testing samples based on an orthogonal table considering load factors, ambient temperature, and wind speed.
Key Insights and Ideas
- Dynamic Hot-Spot Location is Key: The paper highlights that under three-phase unbalanced conditions, the hot-spot location is not static but moves between phases. This necessitates a method that can adapt to these dynamic shifts.
- Streamlines as Thermal Pathways: Analyzing the oil flow streamlines provides a physically intuitive way to understand how heat is transported and to select relevant temperature measurement points on the transformer's exterior that correlate with the internal hot spots.
- Machine Learning for Complex Thermal Dynamics: SVR proves to be an effective tool for capturing the non-linear relationship between external temperature measurements and internal hot-spot temperatures, especially under challenging unbalanced load conditions.
- Genetic Algorithms for Feature Engineering: Using a genetic algorithm to reduce the number of input features (steel tank temperatures) is crucial for making the model practical and efficient without sacrificing accuracy. This demonstrates the power of evolutionary computation in optimizing machine learning models.
Key Results
- The proposed inversion model achieved high accuracy under three-phase unbalanced conditions, with a maximum temperature difference between calculated and inverted HST of 3.71 K.
- The mean absolute percentage error (MAPE) was 1.35%, indicating good overall accuracy.
- After dimensionality reduction using a genetic algorithm, the number of input features was reduced from 24 to 7 (corresponding to 8 dimensions in the SVR model), significantly improving efficiency and operability.
- The mean square error (eMSE) of the SVR model decreased after dimensionality reduction, indicating improved learning ability.
- The model demonstrated good robustness, with a maximum HST inversion difference of 3.64 K even after adding random Gaussian noise to the test samples, suggesting it is less affected by measurement errors.
Limitations and Future Work
- Simulation-Based Data: While validated with experimental data, the primary training dataset is derived from simulations. The accuracy of the model is thus dependent on the fidelity of the thermal fluid field simulation.
- Specific Transformer Model: The research focused on a specific transformer model (S13-M-100 kVA/10 kV). Further validation on different transformer types and sizes would be beneficial.
- Practical Implementation: While the dimensionality reduction improves operability, the paper suggests that further optimization of input temperature measurement points might be needed for even greater practicality.
The authors suggest that their method, based on accurate thermal fluid field simulation and advanced inversion techniques, can be applied to HST inversion in larger transformers as well, provided high-precision thermal fluid field simulations are available.
Summary updated