Comparative Study of Autoencoder and LSTM-AE for Extreme Temperature Anomaly Detection in Semarang

Abstract

Climate change has increased the frequency and intensity of extreme weather events, including heatwaves and cold spells, posing critical risks to public health and urban infrastructure. This study proposes and compares two deep learning frameworks based on Autoencoders, namely the Long Short-Term Memory Autoencoder (LSTM-AE) and the standard Autoencoder (AE), for detecting extreme temperature anomalies using historical daily data from 2005 to 2025 in Semarang City. Unlike conventional anomaly detection methods, the LSTM-AE introduces temporal learning through recurrent memory cells, enabling it to capture sequential temperature dependencies that static AE models cannot. Both models are trained to reconstruct “normal” temperature patterns, with anomalies identified when reconstruction errors exceed the 95th percentile threshold. The results demonstrate that the LSTM-AE more consistently identifies significant heatwave and cold spell events, with seasonal alarm rates that closely align with local climatic transitions. Several detected peaks coincide with historically documented events such as the 2015–2019 El Niño and 2019–2020 transition periods reported by BMKG, confirming climatological relevance. In contrast, the standard AE detects a higher number of anomalies (726 vs 366 from the LSTM AE) but tends to generate false alarms outside transitional periods. Model performance is evaluated using reconstruction error distributions, Jaccard similarity indices, and monthly alarm rates. This study highlights the potential of LSTM-based architectures for improving anomaly detection in climate data and contributes to developing data-driven strategies for urban climate resilience in tropical regions.