Multi-scale Spatio-temporal graph neural network for enhanced water demand forecasting

Area of Science:

• Environmental Engineering
• Data Science
• Artificial Intelligence

Background:

• Accurate Water Demand Forecasting (WDF) is crucial for efficient Water Distribution System (WDS) management.
• Graph neural networks (GNNs) are commonly used for WDF, but existing methods struggle with single time scales and static spatial graphs.
• These limitations hinder performance, especially with complex systems and long-term forecasts.

Purpose of the Study:

• To propose a novel Multi-scale Spatio-Temporal Graph Neural Network (MSTGNN) for enhanced WDF.
• To address limitations of existing GNNs in capturing multi-scale temporal dependencies and adaptive spatial relationships.
• To improve the accuracy and scalability of WDF in complex WDS.

Main Methods:

• Developed MSTGNN to model the hierarchical nature of water demand time series.
• Constructed hierarchical temporal representations from fine to coarse time scales.
• Learned adaptive, scale-specific graph structures to capture dynamic inter-sensor dependencies.

Main Results:

• MSTGNN demonstrated superior performance in day-ahead WDF compared to six state-of-the-art methods.
• Achieved high accuracy in forecasting water demand at 15-minute intervals using a real-world dataset.
• Showcased significant improvements in forecasting accuracy and scalability.

Conclusions:

• MSTGNN effectively models multi-scale spatio-temporal dependencies in WDS.
• The proposed method offers a robust solution for accurate and scalable Water Demand Forecasting.
• Supports the development of advanced smart applications for WDS management.