Deep learning-based detection and identification of power quality in electrical grids
Li, Xinqi (2024)
Kandidaatintutkielma
2024
School of Energy Systems, Sähkötekniikka
Kaikki oikeudet pidätetään.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi-fe2024050727274
https://urn.fi/URN:NBN:fi-fe2024050727274
Tiivistelmä
As power electronic technology advances and distributed generation becomes more prevalent, the grid is increasingly exposed to a variety of intricate power quality disturbances. These disturbances pose significant risks to the safety and efficiency of electricity users, thus placing power quality issues at the forefront of scientific research.
The primary contributions of this thesis include the development of sophisticated models capable of handling both single and composite power quality disturbance signals. The study meticulously examines the origins and risks associated with seven distinct types of single disturbances and five types of composite disturbances. Through simulation, the research generates a comprehensive dataset of these twelve disturbance types, thereby furnishing a robust feature set for subsequent deep learning-based classification and analysis.
The core of the research is anchored in a novel power quality detection framework utilizing deep residual networks. By incorporating residual modules, the study successfully deepens the convolutional neural network architecture, effectively mitigating the issue of network degradation commonly seen in deep classification networks. An extensive exploration into the impact of parameter configurations within these deep residual networks on disturbance classification and identification is conducted. This results in the determination of an optimal network structure tailored for power quality detection. The findings reveal that deep residual networks not only address the degradation problems of existing models but also significantly enhance detection precision, noise immunity, and expedite network convergence, marking a substantial advancement in the field of power quality management.
The primary contributions of this thesis include the development of sophisticated models capable of handling both single and composite power quality disturbance signals. The study meticulously examines the origins and risks associated with seven distinct types of single disturbances and five types of composite disturbances. Through simulation, the research generates a comprehensive dataset of these twelve disturbance types, thereby furnishing a robust feature set for subsequent deep learning-based classification and analysis.
The core of the research is anchored in a novel power quality detection framework utilizing deep residual networks. By incorporating residual modules, the study successfully deepens the convolutional neural network architecture, effectively mitigating the issue of network degradation commonly seen in deep classification networks. An extensive exploration into the impact of parameter configurations within these deep residual networks on disturbance classification and identification is conducted. This results in the determination of an optimal network structure tailored for power quality detection. The findings reveal that deep residual networks not only address the degradation problems of existing models but also significantly enhance detection precision, noise immunity, and expedite network convergence, marking a substantial advancement in the field of power quality management.