Man et al. (2026) Multi-Target Water Demand Forecasting with Graph Neural Networks: A Comparative Study

Identification

• Journal: Water Resources Management
• Year: 2026
• Date: 2026-03-01
• Authors: Yacan Man, Xiao Zhou, Rui Yuan, Kuizu Su, Shuming Liu
• DOI: 10.1007/s11269-026-04501-w

Research Groups

• College of Civil Engineering, Hefei University of Technology, Hefei, China
• School of Environment, Tsinghua University, Beijing, China

Short Summary

This study systematically evaluates Graph Neural Networks (GNNs) for multi-target water demand forecasting (MTF), demonstrating their superior accuracy and robustness compared to traditional sequence-based models. Self-learning GNNs, specifically MTGNN and MTGODE, achieved enhanced accuracy and stability, particularly under data irregularities and for multi-step predictions.

Objective

• To systematically evaluate the performance, interpretability, and robustness of representative Graph Neural Networks (GNNs) for multi-target water demand forecasting (MTF) tasks, comparing them against sequence-based models.
• To investigate the relative importance of auxiliary meteorological and temporal features using SHapley Additive exPlanations (SHAP) analysis and assess their influence on forecasting accuracy.
• To evaluate the robustness of GNNs under critical operational conditions, including sensor outages and heterogeneous monitoring frequencies, commonly encountered in real-world water supply systems.

Study Configuration

• Spatial Scale: 10 hydraulically interconnected districts within a large urban water supply system (WSS).
• Temporal Scale: Data collected from July 1, 2022, to July 30, 2023, with an hourly temporal resolution (9,504 observations over 396 consecutive days). Forecasting horizons included single-step (1 hour) and multi-step (up to 12 hours).

Methodology and Data

• Models used:
  • Graph Neural Networks (GNNs): MTGNN, MTGODE (self-learning GNNs), STGCN, DCRNN (predefined GNNs), StemGNN.
  • Baseline Models: Long Short-Term Memory (LSTM), Transformer-based networks (STformer, STAEformer), Support Vector Regression (SVR).
• Data sources: Historical water demand records from 10 districts of an urban WSS, supplemented with auxiliary meteorological and temporal features.

Main Results

• GNN-based models consistently outperformed sequence-based approaches in single-step water demand forecasting tasks.
• Self-learning GNNs, MTGNN and MTGODE, achieved the best overall performance in single-step forecasting, with mean R² values of 0.960 and 0.962, and Mean Absolute Percentage Error (MAPE) ranges of 6.400–8.322% and 6.236–7.265%, respectively.
• In multi-step forecasting (up to a 12-step horizon), MTGNN and MTGODE maintained relatively stable performance, achieving R² ≥ 0.933 and MAPE ≤ 14.997%, demonstrating stronger robustness compared to baseline models.
• SHAP analysis identified ultraviolet radiation intensity (UVIntensity) and 2 m air temperature (Temp2m) as the most influential meteorological factors for MTF.
• Incorporating temporal features (hour-of-day, day-of-week) and the top three SHAP-ranked meteorological variables (UVIntensity, Temp2m, Humidity) significantly enhanced long-term forecasting accuracy (R² > 0.94).
• Self-learning GNNs (exemplified by MTGODE) demonstrated strong robustness under data irregularities, such as sensor outages and low-frequency sampling, maintaining effective predictive capability (mean MAE of 64.02 m³/h under sensor failures) without the progressive error accumulation observed in recursive prediction strategies like LSTM.

Contributions

• Provided a systematic and comparative evaluation of representative GNN-based models for multi-target water demand forecasting, including both predefined and self-learning architectures, against traditional sequence-based models.
• Established a unified evaluation framework for assessing forecasting behavior across different temporal horizons (single- and multi-step).
• Integrated SHAP-based feature analysis to identify and prioritize influential auxiliary meteorological and temporal features, demonstrating their impact on enhancing forecasting performance, particularly for longer prediction horizons.
• Conducted a comprehensive evaluation of GNN robustness under realistic data irregularities (sensor outages and heterogeneous monitoring frequencies), highlighting their ability to mitigate error accumulation and maintain stable predictive accuracy.
• Offered empirical insights into the applicability of GNN-based approaches for complex MTF tasks and provided practical guidance for model selection and deployment in intelligent water supply systems management.

Funding

• National Natural Science Foundation of China (Grant No. 52370098)
• Joint Fund Project of the Natural Science Foundation of Anhui Province (Grant No. 2308085US05)
• Start-up funding of Hefei University of Technology

Citation

@article{Man2026MultiTarget,
  author = {Man, Yacan and Zhou, Xiao and Yuan, Rui and Su, Kuizu and Liu, Shuming},
  title = {Multi-Target Water Demand Forecasting with Graph Neural Networks: A Comparative Study},
  journal = {Water Resources Management},
  year = {2026},
  doi = {10.1007/s11269-026-04501-w},
  url = {https://doi.org/10.1007/s11269-026-04501-w}
}