CLASSIFICATION OF TIME SERIES DATA BASED ON LATENT MOTIVES FOR EXTREME WEATHER PREDICTION

• Main Author: Aflita Rahmawati, Ratih
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/74541
• Institution: Institut Teknologi Bandung

Extreme weather events have the potential to cause significant damage and disasters in affected regions, highlighting the need for accurate prediction for effective disaster mitigation. This study proposes a latent motif-based time series data classification approach for extreme weather prediction as a mitigation ????!fort. The research involves the development of a system that combines classification algorithms and latent motif discove,y techniques to identify recurring latent patterns that represent the main characteristics of the tilne series data. The results of the system development demonstrate its ability to predict extreme weather events, withf-score pe,jormance of0.97 011 the AWS BMKG dataset and 0. 76 on the Kaggle dataset. Moreover, the developed model can predict extreme weather events 24 hours before their occurrence, achieving a recall rate of 0.94 011 the AWS BMKG dataset, and 36 hours before their occurrence with a recall rate of 0.90 011 the Kaggle dataset. The recall results indicate that the system developed in this research can accurately detect 90-94% of extreme weather events within the 24-36 hour timeframe before the events occw: Through the conducted experimental evaluations, the latent motif-based classification method can predict extreme weather events by leveraging recurring latent patterns in the time series data. Predictions are made by forming a feature space that indicates the occurrence of latent motifs in a time series prior to the onset of extreme Heather events. The occurrence of mot(f'I is represented by the similarity distance between segment o the series and a latent motif,' or by the frequency<????( occuJ'l'ellce of a lat 11/ motif in the entire series. This study conc/11(/es that the developed system shows significant potential for application in tire develop111e11/ r????/' lime series data classification techniques that ca11 be 11/i/izedfor prediction p1111J(JSC ••