%0 Journal Article
%T 基于Vision Transformer和多头注意机制的频率稳定性预测方法<br>Frequency Stability Prediction Method Based on Vision Transformer and Multi-Headed Self-Attention
%A 杜东来
%A 荣娜
%A 张异浩
%J Modeling and Simulation
%P 5354-5367
%@ 2324-870X
%D 2023
%I Hans Publishing
%R 10.12677/MOS.2023.126486
%X 电力系统频率稳定性预测(frequency stability prediction, FSP)对于快速准确地制定故障后控制措施具有重要意义。然而，传统的数据驱动方法未能有效地将系统的时空特征纳入模型训练中，并且存在系统信息利用不足、抗噪能力差和面对新拓扑的泛化能力差等问题。针对上述问题，本文提出了一种基于多头注意机制和视觉变压器(Vision Transformer, ViT)的FSP预测方法。首先，所提出的预测方法利用位置编码层来捕捉系统的拓扑结构，以获得空间信息。其次，通过使用多通道输入层来捕获系统时间序列数据以获得时序信息。然后，采用基于多头注意力机制层来增强ViT模型的泛化能力和鲁棒性。在修改的新英格兰39总线系统上进行了测试。实验结果表明，与传统深度学习方法相比，本文所提ViT模型误差值在一般情况下降低了65.71%；在30 dB、20 dB和10 dB噪声环境下降低了61.55%、56.61%和47.14%；在拓扑改变环境下降低了49.09%，这证明了所提模型具有更高的预测精度、更好的鲁棒性和泛化能力。<br />
Frequency stability prediction (FSP) of power systems is of great significance for quickly and accu-rately developing post fault control measures. However, traditional data-driven methods fail to ef-fectively incorporate the spatiotemporal features of the system into model training, and there are problems such as insufficient utilization of system information, poor noise resistance, and poor generalization ability in the face of new topologies. In response to the above issues, this article pro-poses an FSP prediction method based on multi-head attention mechanism and Vision Transformer (ViT). Firstly, the proposed prediction method utilizes a position encoding layer to capture the top-ological structure of the system and obtain spatial information. Secondly, by using multi-channel input layers to capture system time series data to obtain temporal information. Then, a multi-head attention mechanism layer is used to enhance the generalization ability and robustness of the ViT model, tested on the modified New England 39 bus system. The experimental results show that compared with traditional deep learning methods, the error value of the ViT model proposed in this paper is generally reduced by 65.71%; Reduced by 61.55%, 56.61%, and 47.14% in 30 dB, 20 dB, and 10 dB noise environments; The reduction of 49.09% in the topology changing environment proves that the proposed model has higher prediction accuracy, better robustness, and generaliza-tion ability.
%K 频率稳定性预测，深度学习，视觉变压器模型，多头注意力机制<br>Frequency Stability Prediction
%K Deep Learning
%K Vision Transformer Model
%K Multi-Headed 
Self-Attention
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=75423