Application of deep learning time series models for anomaly detection in hydrological reporting data

Abstract

To further improve the quality management system of hydrological flood reporting data and achieve real-time anomaly monitoring, this study explores the cross-disciplinary application of deep learning algorithms in hydrology. By analyzing the time-series characteristics of hydrological reporting data, a hybrid model combining Convolutional Neural Networks(CNN) and Long Short-Term Memory(LSTM) networks is proposed to enhance feature extraction and processing. Furthermore, considering the physical process of river flow, spatial correlation features embedded in upstream and downstream water level and flow data are integrated to enable multi-step time-series forecasting. A dataset comprising three years of monitoring data from representative stations across the upper, middle, and lower reaches of the Yangtze River was constructed for simulation. The result demonstrate that the proposed CNN-LSTM model achieves a coefficient of determination(R²) of 0.77 for water level prediction and 0.84 for flow simulation, indicating good performance in hydrological forecasting. Building upon this, an integrated anomaly detection framework based on “prediction-residual-discrimination” is developed. In the case study of the Three Gorges Reservoir, the anomaly detection rate reached 100%, with an R² of 0.991 for water level simulation, effectively enabling the identification of abnormal flood reporting data. This study provides a robust intelligent algorithmic foundation for enhancing the quality of hydrological reporting data in river basins.

Keywords

water level prediction / flow prediction / convolutional neural network(CNN) / multi-step time series forecasting / anomaly detection