Generative AI for spatio-temporal multivariate imputation and demand prediction in water distribution systems

Accurate forecasting of water demand is essential for optimizing the management and operation of water distribution networks (WDNs), ensuring sustainable resource utilization and service reliability. The deployment of Internet of Things (IoT) technologies has enabled real-time monitoring of critical parameters such as water pressure and water level. However, sensor failures, communication interruptions, and environmental disturbances often result in multivariate missing data from heterogeneous WDN sensors, posing significant challenges for reliable prediction. Moreover, the complex spatial and temporal interdependencies among sensor measurements further complicate the forecasting process. To address these challenges, this study proposes a novel Spatio-Temporal Imputation Model for Multivariate Water Demand Prediction (STIM). The model integrates spatial correlations and temporal dependencies within multivariate IoT sensor data, employing a generative AI-based imputation mechanism enhanced with Bi-directional Long Short-Term Memory (Bi-LSTM) and Hypergraph Neural Network (HGNN) modules to accurately reconstruct missing data. A deep neural network is then utilized for short-term water demand forecasting based on the imputed dataset. The framework is validated using real-world data collected from an IoT-enabled WDN, incorporating both water level and pressure measurements. The model is validated on real-world WDN data, including both the original monsoon period and an additional summer dataset, demonstrating robustness across seasonal variations. Comprehensive evaluations against seven state-of-the-art AI models demonstrate the superior performance of the proposed STIM model. The model achieves a coefficient of determination ( R 2 ) up to 0.90 for imputation and 0.87 for demand prediction, with mean absolute error (MAE) ranging from 0.031 to 0.071 across varying missing data rates. Even under extreme missingness levels of 90%, the model maintains robust performance, achieving imputation R 2 of 0.87 and prediction MAE of 0.071. Extensive ablation studies further confirm the contribution of each model component, demonstrating robustness, accuracy, and practical applicability of the STIM model. The proposed framework offers a highly reliable decision-support tool for real-time water demand management in IoT-enabled WDNs, even under challenging multivariate missing data scenarios.