TFusionAE:An adaptive framework for meteorological data anomaly detection with Time-Frequency Fusion Autoencoder

Meteorological data plays a pivotal role in areas such as weather forecasting and new energy consumption. However, noise is inevitably introduced during the data collection process, and accurate meteorological data forms the foundation for downstream tasks, making anomaly detection on such data especially important. The complexity and diversity of the data pose significant challenges to detection, with existing methods struggling to balance noise sensitivity and temporal dependency, thus failing to identify all anomalies. Moreover, manual threshold adjustments introduce additional costs. In this paper, we propose a novel anomaly detection framework for meteorological data named Time-Frequency Fusion Autoencoder (TFusionAE). It reduces the impact of noise through Variational Mode Decomposition (VMD) and integrates local numerical features with sequential pattern features to accurately capture anomalies. Additionally, we incorporate the Tyrannosaurus Rex Optimization Algorithm for automatic threshold optimization, which not only enhances the anomaly detection capability but also boosts the adaptability of the framework. Experiments conducted on four common types of meteorological data—temperature, wind speed, air pressure, and humidity—demonstrate the effectiveness and generalization ability of TFusionAE.