A Machine Learning-Based Approach for Link Recovery in Smart Grids Using Software-Defined Networking

Abstract

Network failures can cause serious damage and disrupt communications, which in turn significantly reduces service reliability. Among the various causes of network failure, link failure between network components is one of the most common. For intelligent, autonomous operation without human intervention, a smart grid must have a flexible communication infrastructure capable of rapidly detecting and repairing link failures. A software-defined networking (SDN) approach offers this capability. SDN enables centralized control and decouples the data and control planes, facilitating dynamic management and rapid response to failures. This technology enhances the speed of rerouting and path recovery during link failures by providing centralized and optimized decision-making. Central controllers in SDN architectures can select new paths for data flows in the event of failure, thereby reducing recovery time and packet loss rates. In this study, we design a module for the SDN controller equipped with link failure detection features and implement automatic recovery using the Q-learning algorithm. In smart grid environments, packet loss rates are critical because any data loss can degrade service quality and hinder the reception of essential information required for system control in crisis situations, ultimately reducing the grid's ability to respond to instabilities. The core focus of this research is reducing the packet loss rate, as data loss severely impacts network stability and efficiency. When comparing the proposed method to a reference study, the packet loss rate, recovery time, and algorithm execution time in the German topology decreased by 66.62%, 85.99%, and 91.99%, respectively. Similarly, in the U.S. topology, these metrics were reduced by 90.50%, 76.99%, and 98.99%, respectively. Additionally, compared to the normal state, the packet loss rate was reduced by an average of 67.95% in the German topology and 13.15% in the U.S. topology.