Water quality anomaly detection research based on GRU-PINN model

Abstract

Maintaining high-quality water resources is essential for sustainable urban water resource management and public health. This paper introduces the GRU-PINN model, developed based on the Gated Recurrent Unit (GRU) network and integrated with a Physics-Informed Neural Network (PINN), to analyze real-world monitoring data from a water treatment company. By embedding domain-specific physical constraints into the loss function, the model enhances interpretability and reduces false alarms. The proposed approach begins with feature engineering on the raw dataset, including missing value imputation, normalization, trend feature extraction, and rolling feature computation. Feature selection is then performed based on feature importance ranking and mutual information analysis. The GRU-PINN model is subsequently employed for anomaly detection in the dataset. The performance of the model is evaluated using the F1-score, precision, and recall. The F1-score, which represents the harmonic mean of precision and recall, is particularly suitable for imbalanced datasets, as the dataset used in this paper contains very few anomaly instances. Experimental results demonstrate that the proposed GRU-PINN model outperforms traditional models by achieving a higher F1-score, thereby improving anomaly detection performance.