Seminario | “An unsupervised deep anomaly detection model combined with extreme value theory: Detecting sudden stratospheric warming events”

Seminario organizado por el grupo de Termodinámica y Computación Cuántica, a cargo de Khulan Myanmar (Universidad de  Mongolia). El seminario es accesible para los estudiantes de grado, a los que se recomienda que asistan.

Resumen:  Sudden stratospheric warming (SSW) events represent intense meteorological phenomena characterized by rapid and significant temperature increases within the polar stratosphere. These events can have a considerable impact on extreme weather conditions in mid-latitude countries. However, research on the estimation and detection of SSW remains limited. This study aims to develop a novel method for detecting SSW events using unsupervised deep anomaly detection algorithms. By leveraging unsupervised learning, patterns and structures in data can be rec- ognized, facilitating the development of predictive models without prelabelled datasets. The long short-term memory-autoencoder (LSTM-AE) with correlation weighting function model were used for the predicting and extreme value distributions (EVD) were applied for detecting anomaly. National Centres for Environmental Prediction (NCEP) and the National Centre for Atmospheric Research (NCAR) reanalysis1 temperature and zonal wind data were used for model training, whereas ERA5 reanalysis data were used for validation. This study focuses on the Northern Hemisphere (60°N-90°N) at the 10 hPa pressure level from 1979–2023. The results of this study indicate that the LSTM-AE-peaks over threshold (LSTM-AE-POT) hybrid model with dynamic correlation weighting loss function is the most effective for accurately predicting separate and joint fluctuations to detect the SSW events. Performance evaluation is conducted using confusion matrix-based metrics and temporal consistency metrics. The results demonstrate that the LSTM-AE-POT hybrid model identifies long-term, persistent anomalies with high accuracy and consistency. Additionally, the results of the validation confirm the model’s stability and smooth detection performance for SSW events.

• Fecha: viernes 13 de junio de 2025
• Lugar: Laboratorio de Fïsica Computacional, Dept de Electromagnetismo y Física de al Materia, Planta Baja, Facultad de Ciencias (junto al péndulo gigante)
• Horario: 12:00 h.
• Organiza: grupo de Termodinámica y Computación Cuántica
• Más información: Michalis Skotiniotis | mskotiniotis@onsager.ugr.es