An automatic algorithm of identifying vulnerable spots of internet data center power systems based on reinforcement learning

Abstract The internet data center (IDC) power system provides power guarantee for cloud computing and other information services, so its importance is self-evident. However, the occurrence time of malignant destructive events such as lightning strikes, errors in operation and cyber-attacks is unpredictable. But the loss can be minimized by formulating coping strategies in advance. So, identifying the vulnerable spots of the IDC power system come to be the key to guarantee the normal operation of information systems. Generally, the IDC power network can be modelled as a graph G, and then, the methods of finding nodes’ centrality can be applied to analyse the vulnerability. By our experience, it is not the best approach. Unlike the previous approaches, we do not solve the issue as the traditional graph problem. Instead, we fully utilize the characteristics of the IDC power network and apply reinforcement learning techniques to identify the vulnerability of the IDC power network. To our best knowledge, it is the first applying of artificial intelligence in traditional IDC power network. In this article, we propose PFEM, a parallel fault evolution model for the IDC power network, which can accelerate the process of electrical fault evolution. Moreover, we designed an algorithm which can automatically find the vulnerable spots of the IDC power network. The experiment on a real IDC power network demonstrate that the impact of vulnerable devices derived from our proposed algorithm after failure is about 5% higher than that of other algorithms, and tripping single-digit electrical devices of the IDC power system with our proposed algorithm will lead to loss of all loads.