Anomaly Detection Model for Hydrological Time Series Based on Spatiotemporal Features

To address the challenge of pattern anomaly detection in hydrological time series, a novel anomaly detection method Temporal and Spatial Feature Fusion based on GCN and TCN (TSFGN) that integrates spatiotemporal features and deep learning is proposed. The method first constructs a prediction model based on deep learning to calculate reconstruction and prediction losses to obtain the sequence anomaly score. Then, it employs the Peak Over Threshold (POT) algorithm to partition the anomaly score subsequences and proposes a Multidimensional Symbolic Aggregate Approximation(MSAX) method to symbolically represent those subsequences. Finally, the top-k subsequences with the highest Average non-self-match distance(Disti) are identified as pattern anomalies. Experimental results on the measured dataset from the Qinhuai River Basin demonstrate that the proposed method can effectively capture anomalous patterns associated with extreme hydrological events, achieving a PR-AUC of 0.9768. This performance surpasses that of existing mainstream methods in both detection accuracy and robustness, indicating a strong precision-recall balance. The proposed approach provides a reliable technical solution for hydrological data quality control and disaster early warning.