Abstract
Traditionally, the ampacity of an overhead transmission line (OHTL) is a static value based on adverse weather conditions, which constrains the transmission capacity. This paper proposes a dynamic ampacity prediction method using a Back-Propagation (BP) neural network optimized by Bayesian optimization (BO-BP). A two-stage current-stepping model is established to provide initial predictions, followed by a risk control strategy that corrects results based on model error quantiles. Simulation results on a 220-kV line demonstrate a 20% improvement in prediction accuracy over traditional heat balance equations and a potential 150% increase in transmission capacity compared to static ratings.
Introduction
As power system loads grow and renewable energy integration increases, the limitations of static line ratings (SLR) become more apparent. SLR relies on "worst-case" weather scenarios—typically high ambient temperatures and low wind speeds—leading to significant underutilization of existing infrastructure. Dynamic Line Rating (DLR) offers a solution by adjusting capacity based on real-time conditions. However, traditional physics-based models are difficult to parameterize accurately due to varying environmental factors, and standard machine learning models often lack reliability when operating outside their training data. This research introduces a hybrid BO-BP approach to bridge this gap, ensuring both high accuracy and physical reliability.
Methodology
BO-BP Neural Network
The core of the proposed method is a BP neural network that models the complex relationship between weather variables and conductor temperature. The model takes five primary features as input: wind speed, wind direction, solar irradiance, ambient temperature, and wire current. To optimize performance, Bayesian Optimization (BO) is used to automatically tune hyperparameters, specifically the learning rate and the number of hidden layer neurons. This process uses Gaussian Process Regression and an acquisition function to efficiently search the parameter space, avoiding the pitfalls of manual tuning and ensuring the model generalizes well to new data.
Two-Stage Current-Stepping Model
To ensure physical consistency and safety, the authors developed a two-stage workflow. In the first stage, the BO-BP model predicts temperature for current ranges within the training set. If the required current exceeds the training data, the system transitions to a second stage using a corrected version of the traditional heat balance equation. This hybrid approach leverages the high accuracy of machine learning for common operating conditions while maintaining the physical bounds of traditional engineering for extreme scenarios.
Risk Control Strategy
To mitigate the risk of conductor overheating due to prediction errors, a risk control strategy is implemented. Instead of using raw model outputs, predictions are adjusted based on the lower $\alpha$-quantile of the historical error distribution. By shifting the current-temperature (I-T) curve, this statistical safety margin allows operators to quantitatively tune the balance between capacity gains and operational risk.
Key Results
The BO-BP model was tested using real-world operational data from a 220-kV transmission line in China. Data from 2022 was used for training, while 2023 data was used for evaluation. Key findings include:
- Enhanced Accuracy: The model achieved a reduction in Root Mean Square Error (RMSE) of over 20% compared to traditional heat balance equations across all seasons. The highest improvements were observed in winter (36.41% RMSE reduction) and fall (31.34%).
- Capacity Gains: The dynamic method improved transmission capacity by more than 150% relative to the static rating of 581 A. Under favorable weather conditions, the available capacity was significantly higher than traditional estimates.
- Safety and Reliability: Even with conservative risk control settings ($\alpha = 1$), capacity improvements remained above 154%, demonstrating that significant efficiency gains are possible without compromising safety standards.
Conclusion
The integration of Bayesian-optimized neural networks with physics-based modeling provides a robust framework for dynamic line rating. By accounting for model errors through a statistical risk control strategy, this approach offers a safe and highly effective way to unlock hidden capacity in modern power grids. Future work will explore online training to adapt to environmental drifts and investigate the long-term impact of higher dynamic ratings on conductor aging and cyber-physical security.
Summary updated