Prediction-based Real-time Anomaly Detection for Water Quality Time Sequences

Real-time anomaly detection is critical to effective and efficient water resource management. Existing methods are either improper for real-time use, or not cost-efficient due to requirements of total verification on all events. In this paper, we present a prediction-based real-time anomaly detection scheme for water quality time sequences. Using a deep learning prediction model, the scheme predicts a new value based on historical data and computes the residual between the predicted value and the next measurement. The scheme then uses Isolation Forest to assign an anomaly score to each residual. The scheme also maintains a threshold. Residuals that exceed the threshold are reported to human resources for verification. The scheme uses the verification results to update the threshold in order to improve detection accuracy for future anomalies. Experimental results show that the proposed scheme achieves high precision and recall rates, especially when the anomaly intensity is high or multiple water quality parameters show abnormal values simultaneously, while significantly reducing false alarms, compared with the scheme which detects anomalies directly based on actual measurements.