Deep Learning-Based Digital Twin Framework for Real-Time Water Network Anomaly Detection

Urban water distribution systems are critical infrastructures that require continuous monitoring to prevent water loss caused by leaks and operational anomalies. Conventional monitoring approaches often struggle to detect early-stage anomalies due to the complexity and temporal dynamics of multivariate sensor data. This study proposes a Deep Learning based Digital Twin framework for real time anomaly detection in urban water networks. The framework integrates a Deep LSTM Autoencoder with a Digital Twin architecture to model normal operational behavior and identify deviations through reconstruction error analysis. The model was trained exclusively on normal operational data and optimized using a Receiver Operating Characteristic based threshold selection strategy. Experimental evaluation on time series sensor data demonstrated that the proposed approach achieved a recall of 0.74 for leak detection and an Area Under the Curve value of 0.750, indicating reliable discrimination between normal and anomalous conditions. The results show that temporal sequence learning significantly enhances anomaly detection capability in complex water distribution systems. By synchronizing real time anomaly detection with a Digital Twin environment, the proposed framework enables proactive monitoring, improved situational awareness, and predictive maintenance support for smart city water management applications.