The Power System Analysis Based on Neural Network

Yang ChenAdvances 2015 · DOI

TL;DR

This paper proposes a novel approach to power system analysis by combining deep neural networks (DNNs) with traditional power system modeling techniques. By leveraging the pattern recognition capabilities of DNNs, the authors aim to improve the accuracy and robustness of modeling generator leading phase operation, a critical aspect of voltage regulation and power quality.

What Problem Does This Paper Address?

The paper addresses the challenge of accurately modeling the complex, nonlinear behavior of synchronous generators, particularly during leading phase operation. Traditional methods often struggle with the inherent nonlinearity, randomness, and interference present in power systems, leading to limited applicability and accuracy. The authors highlight the need for more sophisticated modeling techniques that can capture these complexities, especially as power systems become more dynamic and integrated with renewable energy sources.

Methods

The core of the proposed approach involves using a deep neural network, specifically a Convolutional Neural Network (CNN) architecture, for power system analysis. The authors draw inspiration from auditory and physiological models for filter bank design, arguing that these perceptually-driven filters can outperform statistically designed ones in certain modeling contexts. The neural network is trained on power system data, with the input being the normalized log of the power spectrum. The paper details the forward and backward propagation mechanisms of the BP neural network and presents a general structure of the proposed neural network, including various layers like convolutional, pooling, normalization, and fully connected layers. They also discuss the use of genetic algorithms to optimize initial weights and thresholds for the neural network, enhancing its precision. Normalization of input samples is also employed to speed up training and prevent saturation.

Datasets

Not explicitly detailed in the paper. The authors mention using "data samples" for training and simulation, implying real-world or simulated power system data, but specific datasets are not listed.

Key Insights and Ideas

  • Perceptual Filter Banks for Power Systems: The paper posits that filter banks designed based on human perception (like those used in auditory processing) can be more effective for power system modeling than purely statistical approaches.
  • DNNs for Complex Power System Dynamics: Deep neural networks, particularly CNNs, are well-suited to capture the intricate nonlinearities and dynamic behaviors of power systems, offering a more robust alternative to traditional methods.
  • Genetic Algorithms for Optimization: Integrating genetic algorithms to optimize neural network parameters (weights and thresholds) can significantly improve the network's accuracy and performance.
  • Generalization is Key: The authors emphasize that while increasing the number of hidden layer neurons can reduce error, excessive neurons can lead to generalization issues, highlighting the importance of finding an optimal network architecture.

Key Results

The paper presents numerical and demonstrative results showing that increasing the number of hidden layer neurons in the proposed neural network generally leads to a decrease in convergence error. However, they observed that beyond a certain point (14-18 neurons), generalization error started to increase. This finding is attributed to potential pathological issues in the output layer weights matrix. The simulation results, visualized in Figure 2, demonstrate the feasibility and effectiveness of the proposed approach in capturing underlying data distributions and formulating hyperbox structures.

Limitations and Future Work

The paper acknowledges that a filter bank designed from perceptual evidence is not always guaranteed to be the best in a statistical modeling framework. A limitation is the lack of specific details regarding the datasets used for training and evaluation. Future work involves conducting more research to leverage deep neural networks and more complex power systems for improved results. The authors also suggest further investigation into the optimal number of hidden layer neurons to avoid generalization errors.

Summary updated

Abstract

With the rapid and bursting development of pattern recognition technique and power system theory, the combination of the two is urgently needed. We argue that a filter bank that is designed from perceptual evidence is not always guaranteed to be the best filter bank in a statistical modeling framework where the end goal is word error rate. Therefore, to combine the proposed techniques with the neural network, we conduct theoretical analysis and numerical simulation on the power systems. The experimental analysis and numerical simulation indicate that our proposed approach performs well and contains satisfactory robustness.
BibTeX
@inproceedings{chen2015power,
  author = {Yang Chen},
  title = {{The Power System Analysis Based on Neural Network}},
  booktitle = {Advances},
  year = {2015},
  doi = {10.2991/ipemec-15.2015.225},
}
Figure