DC-Approximated Power System Reliability Predictions with Graph Convolutional Neural Networks

Abstract

The current standard operational strategy within electrical power systems is done following deterministic reliability practices. These practices are deemed to be secure under most operating situations when considering power system security, but as the deterministic practices do not consider the probability and consequences of operation, the operating situation may often become either too strict or not strict enough. This can in periods lead to inefficient operation when regarding the socio-economic aspects. With the continuous integration of renewable energy sources to the electrical power system coupled with the increasing demand for electricity, the power systems have been pushed to operating closer to their stability limit. This poses a challenge for the operation and planning of the power system. Research is therefore being invested into finding more flexible operational strategies which operates according to probabilistic reliability criteria, taking the probability of future events into consideration while also aiming to minimize the expected cost and defining limits for probabilistic reliability indicators.

To reliably plan and operate the systems according to a probabilistic reliability criterion, numerical problems such as the Optimal Power Flow (OPF) and the Power Flow (PF) equations are used. These tools are helpful as they are used to determine the optimal way of producing and transporting power. These tools are also used in contingency analyses, where the effect of occurring contingencies is analyzed and evaluated. Due to the non-linearity of the PF equations, the solution is often found through iterative numerical methods such as the Gauss-Seidel method or the Newton-Raphson method. These numerical methods are often computationally expensive, and convergence to the global minimum is not guaranteed either. In recent years, various Machine Learning (ML) models have gathered a lot of attention due to their success in different numerical tasks, particularly Graph Convolutional Networks (GCNs) due to the model’s ability to utilize the topology and learn localized features. As the field of GCN is new, extensive research is being committed to identify the GCNs ability to work on applications such as the electrical power system.

This thesis seeks to conduct preliminary experiments where Graph Convolutional Networks (GCN) models are used as a substitution for the numerical DC-OPFs which are used to determine values such as the system load shedding due to contingencies. The GCN models are trained and tested on multiple datasets on both a system- and a node-level, where the goal is to test the models' ability to generalize across perturbations of different system-parameters, such as the system load, the number of induced contingencies and different system topologies.

The experiments of the thesis show that the GCNs can predict the load-shedding values across multiple system-parameter perturbations such as the number of induced contingencies, increasing load-variation and a modified system-topology with a high accuracy, without having to be retrained for those specific situations. Though, the further the system-parameters were perturbated, the less accurate the model's predictions became. This reduction in accuracy per system-parameter perturbation was caused by a change in the load-shedding pattern as additional parameters were perturbated, which the models were unable to comprehend. Lastly, this thesis also shows that the GCN models are substantially faster than the numerical methods which they seek to replace.