<p>With the acceleration of urbanization, short-term water demand forecasting is crucial for the efficient management of water distribution networks (WDNs). However, current methods that use water node data as feature vectors to construct spatial‒temporal graphs for water demand forecasting have limitations. Existing methods using graph convolutional networks (GCNs) primarily capture global correlations between node feature vectors when extracting spatial features, neglecting dynamic correlations at different times. Additionally, they fail to fully consider the multiscale temporal features of water demand data when extracting temporal features, which reduces the accuracy of predictions. To address these issues, this study proposes a novel multitype graph fusion framework (MGFF) for short-term water demand prediction, which includes graph feature extraction, graph feature fusion, and output modules. The framework first constructs three types of graph data—spatial, temporal, and mixed-scale graphs—by adjusting the node embedding dimensions. The GCN and the designed recurrent module are then used to extract the corresponding feature representations, capturing global, dynamic, and multiscale correlations. In the graph feature fusion module, a stepwise feature fusion mechanism is subsequently designed to integrate the multitype graph features, capturing the dynamically interactive spatial‒temporal dependencies in the water demand data and thereby improving the prediction accuracy. Lastly, the framework incorporates an output module composed of convolutional layers, residual mechanisms, and fully connected layers to produce predictions. Experiments on real-world WDNs demonstrate that the proposed framework achieves greater accuracy than do models such as long short-term memory (LSTM) and spatial–temporal fusion graph neural networks (STFGNNs).</p> Graphical Abstract <p></p>