Abstract
A massive challenge for future cities is being environmentally sustainable by incorporating renewable energy resources (RES). At the same time, future smart cities need to support resilient environments against cyberthreats on their supported information and communication technologies (ICT). Therefore, the cybersecurity of future smart cities and their smart grids is of paramount importance, especially on how to detect cyber-attacks with growing uncertainties, such as frequent topological changes and RES of intermittent nature. Such raised uncertainties can cause a significant change in the underlying distribution of measurements and system states. In such an environment, historical measured data will not accurately exhibit the current network's operating point. Hence, future power grids' dynamic behaviors within smart cities are much more complicated than the conventional ones, leading to incorrect classification of the new instances by the current attack detectors. In this paper, to address this problem, a long short-term memory (LSTM) recurrent neural network (RNN) is carefully designed by embedding the dynamically time-evolving power system's characteristics to accurately model the dynamic behaviors of modern power grids that are influenced by RES or system reconfiguration to distinguish natural smart grid changes and real-time attacks. The proposed framework's performance is evaluated using the IEEE 14-bus system using real-world load data with multiple attack cases such as attacks to the network after a line outage and combination of RES. Results confirm that the developed LSTM-based attack detection model has a generalization ability to catch modern power grids' dynamic behaviors, excelling current traditional approaches in the designed case studies and achieves accuracy higher than 90% in all experiments.