Graph Machine Learning Models Applied on the Identification of Critical Transmission Lines in Electrical Power Systems
Name: ROGERIO JOSÉ MENEZES ALVES
Publication date: 04/04/2022
Advisor:
Name |
Role |
|---|---|
| MARCIA HELENA MOREIRA PAIVA | Advisor * |
Examining board:
| Name |
Role |
|---|---|
| JORGE LEONID ACHING SAMATELO | External Examiner * |
| MARCIA HELENA MOREIRA PAIVA | Advisor * |
Summary: When the operation of electric power systems is concerned, one may associate security with predicting a future event, and then be always ready to what may happen in the future. Obtaining a more secure system becomes a task deeply entangled with having robust and solid contingency plans and operators are required to continuously evaluate their contingency strategies to ensure the fastest response to a contingency scenario that might not have happened before. The number of possible events increases even further when not
only single contingencies are evaluated, but also multiple contingencies. Therefore, one must seek for the balance between desired security level and practical cost-efectiveness. The initial step of evaluating the scenarios that will be studied and simulated in detail is called the screening step, which is critical to the contingency analysis and security assessment of power systems, mainly because it constrains the future study scenarios on a limited set of contingencies. A critical event that is not mapped in the screening step will not be analyzed in further simulations, and might represent an issue in the systems event response mechanisms. On the other hand, the screening is vital because the number of possible contingency scenarios turns impractical to make an exhaustive detailed simulation approach. One alternative approach to make analyses on power systems is to consider topological aspects instead of, or even together with electrical information from the systems. In this case, graph models of the power systems can be constructed and evaluated. Part of the contingency analysis process consists on evaluating the possibly most critical contigencies for a power system, which can be framed as a binary (or possibly multi-class) classification problem.Iin this dissertation, Graph Machine Learning techniques are evaluated on graph models constructed from the system data. Three approaches for learning architectures are proposed and applied to a set of test systems that are commonly used in benchmarks. The
learning approaches are trained for reproducing a known criticality index for transmission lines in a graph model, and the obtained results are analyzed.
