Icing Forecasting of Transmission Lines with a Modified Back Propagation Neural Network-Support Vector Machine-Extreme Learning Machine with Kernel (BPNN-SVM-KELM) Based on the Variance-Covariance Weight Determination Method

Dongxiao Niu, Yi Liang, Haichao Wang, Meng Wang, Wei‐Chiang HongEnergies 2017 · DOI

TL;DR

This paper introduces a novel hybrid model (BPNN-SVM-KELM) for predicting transmission line icing thickness, combining optimized versions of Back Propagation Neural Networks (BPNN), Support Vector Machines (SVM), and Extreme Learning Machines with Kernels (KELM). By leveraging variance-covariance weighting, this ensemble approach significantly improves both the accuracy and stability of icing forecasts, offering crucial technical insights for power grid safety management.

What Problem Does This Paper Address?

Transmission line icing poses a severe threat to power grid reliability, leading to catastrophic failures like tower collapses and widespread power outages. Accurate forecasting of icing thickness is therefore critical for proactive maintenance and preventing these costly disruptions. Existing methods, whether traditional physical or statistical models, often struggle with the inherent uncertainties and complex factors influencing icing. While modern intelligent models like BPNN and SVM show promise, they can suffer from issues like slow convergence, falling into local optima, or difficulties with large datasets. This paper aims to bridge this gap by developing a more robust and accurate combined forecasting model that overcomes the limitations of individual intelligent techniques.

Methods

The core of this research lies in a multi-stage approach to build a superior icing forecasting model:

  • Optimized BPNN (MEC-BPNN): The standard BPNN's tendency to get stuck in local optima and converge slowly is addressed by using a Mind Evolutionary Computation (MEC) algorithm. MEC, with its unique convergence and dissimilation operators, is employed to optimize the initial weights and thresholds of the BPNN, leading to faster and more reliable training.
  • Optimized SVM (BA-SVM): Support Vector Machines (SVM) are powerful for non-linear problems, but their performance heavily relies on selecting optimal parameters (penalty parameter 'c', kernel parameter 'g', and epsilon 'p'). This paper uses the Bat Algorithm (BA), a meta-heuristic inspired by bat echolocation, to efficiently search for and optimize these crucial SVM parameters, enhancing its prediction accuracy.
  • Kernel Extreme Learning Machine (KELM): To further boost regression prediction accuracy and stability, an Extreme Learning Machine with Kernel (KELM) is utilized. KELM, an extension of ELM, incorporates kernel functions to map data into higher-dimensional spaces, improving generalization ability and overcoming the stability issues sometimes seen in basic ELM.
  • Variance-Covariance (VC) Weight Determination: The predictions from the three optimized models (MEC-BPNN, BA-SVM, KELM) are not simply averaged. Instead, a Variance-Covariance (VC) combined method is employed to dynamically determine the weights for each model's prediction. This method considers the variance of each model's prediction errors, allowing the combination to adaptively lean towards the models that perform better on the given data, thus improving overall robustness and accuracy.

Datasets

The study utilizes real-world data from the "Dong-Chao line" and "Tianshang line" in Hunan, China, which experienced severe icing events in 2008 and 2007-2008, respectively. The datasets include:

  • Icing thickness
  • Temperature
  • Relative air humidity
  • Wind speed
  • Wind direction

For each case study, the data was collected at a 2-hour frequency. A portion of the data (e.g., the first 192 samples for the "Dong-Chao line") was used for training, and the remaining data (e.g., the latter 120 samples) was used for testing.

Key Insights and Ideas

  • Ensemble Power: Combining multiple specialized models, each optimized to address specific weaknesses, is more effective than relying on a single model. The VC weighting scheme is crucial for this synergy.
  • Dynamic Weighting is Key: The VC method's ability to dynamically adjust weights based on model performance is a significant advantage. This allows the combined model to adapt to different data patterns and improve robustness, as demonstrated by the differing optimal weights between the two case studies.
  • Addressing Individual Model Weaknesses: The paper effectively demonstrates how advanced optimization techniques (MEC for BPNN, BA for SVM) and architectural enhancements (KELM) can mitigate the inherent limitations of their base models.
  • Wind Direction Complexity: The paper acknowledges the significant impact of wind direction on icing and incorporates a data preprocessing step (clustering) to handle its complex influence, which is vital for accurate forecasting.

Key Results

The proposed BPNN-SVM-KELM combined model consistently outperformed individual models and basic BPNN/SVM in both case studies.

  • Accuracy: The combined model achieved the lowest RMSE, MAPE, and AAE values across both case studies. For instance, in the first case study, the RMSE of the combined model was 0.86%, significantly lower than BPNN (4.44%), SVM (3.68%), MEC-BPNN (1.04%), BA-SVM (1.32%), and KELM (1.32%).
  • Stability: While the individual improved models (MEC-BPNN, BA-SVM, KELM) showed better accuracy than the basic BPNN and SVM, their stability was still lacking. The combined model, however, demonstrated superior stability, with its forecasting curves closely tracking the actual data and exhibiting much smaller fluctuation ranges.
  • Weight Distribution: The VC method assigned weights that leaned towards the best-performing individual model for a given dataset, highlighting its adaptive nature. For example, in Case 1, the weights were 0.42 (MEC-BPNN), 0.34 (BA-SVM), and 0.24 (KELM), while in Case 2, they shifted to 0.40 (MEC-BPNN), 0.25 (BA-SVM), and 0.35 (KELM).

Limitations and Future Work

  • Data Specificity: The models were trained and tested on specific datasets from Hunan, China. While the results are promising, their generalizability to other geographical regions or different icing conditions might require further validation.
  • Computational Cost: While the individual optimized models are faster than traditional methods, the ensemble approach, especially with optimization algorithms, can still be computationally intensive.
  • Future Work: The authors suggest that this research provides a useful reference for the power sector. Future work could explore incorporating more meteorological factors, real-time sensor data, or investigating the model's performance under extreme weather events. Further research into more advanced ensemble techniques or different optimization algorithms could also be beneficial.

Summary updated

Abstract

Stable and accurate forecasting of icing thickness is of great significance for the safe operation of the power grid. In order to improve the robustness and accuracy of such forecasting, this paper proposes an innovative combination forecasting model using a modified Back Propagation Neural Network-Support Vector Machine-Extreme Learning Machine with Kernel (BPNN-SVM-KELM) based on the variance-covariance (VC) weight determination method. Firstly, the initial weights and thresholds of BPNN are optimized by mind evolutionary computation (MEC) to prevent the BPNN from falling into local optima and speed up its convergence. Secondly, a bat algorithm (BA) is utilized to optimize the key parameters of SVM. Thirdly, the kernel function is introduced into an extreme learning machine (ELM) to improve the regression prediction accuracy of the model. Lastly, after adopting the above three modified models to predict, the variance-covariance weight determination method is applied to combine the forecasting results. Through performance verification of the model by real-world examples, the results show that the forecasting accuracy of the three individual modified models proposed in this paper has been improved, but the stability is poor, whereas the combination forecasting method proposed in this paper is not only accurate, but also stable. As a result, it can provide technical reference for the safety management of power grid.
BibTeX
@inproceedings{niu2017icing,
  author = {Dongxiao Niu and Yi Liang and Haichao Wang and Meng Wang and Wei‐Chiang Hong},
  title = {{Icing Forecasting of Transmission Lines with a Modified Back Propagation Neural Network-Support Vector Machine-Extreme Learning Machine with Kernel (BPNN-SVM-KELM) Based on the Variance-Covariance Weight Determination Method}},
  booktitle = {Energies},
  year = {2017},
  doi = {10.3390/en10081196},
}
Figure