In recent years, the size and complexity of Electrical Power and Energy Systems (EPES) has grown significantly, inevitably increasing the occurrences of widespread failures that have led to blackouts on the process. Due to the detrimental socioeconomic impact of such events, there is an increasing need for efficient control strategies, able to protect the grid from propagating failures and alleviate the impact of these contingencies. By integrating micro-grid architectures inside the modern EPES, system operators are able to design new techniques for enhancing the resilience and reliability of the electrical grid.
Micro-grids enable the incorporation of distributed energy resources (DERs) in the system, which typically consist of renewable energy sources that offer several modular and versatile features and can be coordinated in a decentralised manner. Micro-grids are characterised by their ability to operate connected to the main grid, but also disconnected in an autonomous islanded mode. Given their flexible nature, they have been employed to increase the stability and resilience of electrical grids in case of emergencies, through the method of intentional islanding.
During intentional islanding the grid is partitioned into several isolated, self-sustained segments, called islands, aiming to prevent further failures when the system is evolving though a cascading propagation stage. In essence, the intentional islanding problem involves the decision of which transmission lines should be disconnected, so that stable islands are formed, and there is minimum loss of supply on the process. Most current solutions rely on optimisation methods to solve the problem, however these cannot always achieve a polynomial time while satisfying all the system constraints. For this reason, we wanted to make use of the powerful fitting and generalisation capabilities offered by deep learning (DL) and provide a highly efficient solution to the problem of intentional islanding using a Graph Convolutional Network (GCN) architecture.
Our approach is thoroughly described in our recently published journal paper  and relies on a GCN architecture to govern the splitting strategy. In order to utilise the GCN layers in our DL model, we first translate the electrical grid into a graph representation, where each vertex in the graph denotes a bus in the system, and each edge indicates a transmission line or transformer. The edges are represented by weighted lines, depicting either information about the active power flow or the adjacency. The required data are extracted by executing a power flow analysis. In case of buses they involve angle, voltage magnitude, and power demand, while for lines they involve active, reactive power flow, and adjacency information. The overall pipeline of our method is depicted in Figure 1.
The main platform that we used to implement our solution is PyTorch , which allows us to execute the algorithm on both powerful GPUs and conventional CPUs as well. For the power system simulation, we used the pandapower  library which offers several useful functions such as powerflow and state estimation, along with a large number of test grids, used to evaluate the performance of our tool. In order to provide a visual insight regarding the developed islanding schemes, we also designed a dashboard that directly communicates with the tool and displays both graphical and numerical results, as shown in Figure 2.
 Sun, Z.; Spyridis, Y.; Lagkas, T.; Sesis, A.; Efstathopoulos, G.; Sarigiannidis, P. End-to-End Deep Graph Convolutional Neural Network Approach for Intentional Islanding in Power Systems Considering Load-Generation Balance. Sensors 2021, 21, 1650. https://doi.org/10.3390/s21051650
 Pytorch.org. 2021. PyTorch. [online] Available at: <https://pytorch.org/>
 “Pandapower”, pandapower, 2021. [Online]. Available: http://www.pandapower.org/