Analysis of VAE-LSTM Performance in Detecting Anomalies in Average Daily Temperature Data in Jakarta 2000-2023

Abstract

Climate change is happening worldwide, so global climate conditions are a major concern. In densely populated urban areas such as Jakarta, it is impossible to avoid the impacts of climate change, particularly the daily changes in air temperature. Therefore, a sophisticated and efficient approach is needed to find inconsistencies in daily air temperature data to provide critical information for sustainable urban planning and efforts to reduce risks. This research will combine two innovative approaches for hybrid anomaly detection. The method combines generative methods and can extract complex features, such as variational autoencoder (VAE), along with the temporal coding capabilities of long-short-term memory (LSTM), a type of Recurrent Neural Network (RNN). The data used in this study is the average daily air temperature data in Jakarta, obtained from the Kemayoran Meteorological Station and provide by the Meteorology, Climatology, and Geophysics Agency (BMKG). The data used is daily from April 2000 to December 2023. The threshold used to detect anomalies was 229.5, which resulted in excellent performance, namely F1-Score 0.985, Recall 1.000, and Precision 0.971. The VAE-LSTM model identified all dates with significant temperature anomalies, including January 21, 2014, February 22, 2014, November 12, 2014, and February 9, 2015. These dates are significant as they represent extreme weather events that can have severe implications for urban planning and climate change adaptation. The anomalies fall into the categories of point and contextual anomalies. This study contributes to climate research by providing evidence of the effectiveness of deep learning-based hybrid models in detecting complex and context-sensitive temperature anomalies.