Voltage Condition Monitoring Method of Accelerator Distribution Network Based on Deep Learning

Abstract

The voltage change process of distribution network of heavy-ion accelerator is complicated, and the condition monitoring method based on a fixed threshold has great limitations. Therefore, a condition monitoring method based on auto-encoder and bidirectional long short-term memory network is proposed. Firstly, the model has the ability to extract the cross correlation, temporal correlation and dependence of multi-dimensional temporal data, the normal monitoring data of distribution network are reconstructed to obtain the reconstruction error. Then, the mahalanobis distance of reconstruction error is calculated as the condition indicator of distribution network, and the probability density distribution of condition indicator is fitted by kernel density estimation method to determine the abnormal threshold of condition indicator. Finally, the contribution degree of each variable is calculated to determine the variables most related to the abnormal changes, so as to achieve the purpose of voltage condition monitoring of distribution network. The results show that the proposed method can detect abnormal changes and trends in monitoring data, so as to accurately and deeply grasp the condition of accelerator distribution network, which is of great significance for implementing machine protection and optimizing power quality of high-power and high-current heavy-ion accelerator in the future.