A graph convolutional neural network water level prediction method based on geographical optimal similarity
By constructing an adjacency matrix based on Geographic Optimal Similarity (GOS) and combining it with a Graph Convolutional Neural Network (GCN), the problem of insufficient spatial correlation capture in traditional water level prediction is solved, achieving higher accuracy and robustness in water level prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGXI ACAD OF ECO-ENVIRONMENTAL SCI & PLANNING
- Filing Date
- 2026-05-19
- Publication Date
- 2026-07-10
AI Technical Summary
Existing technologies struggle to effectively capture the complex nonlinear spatial correlations between hydrological stations. Traditional adjacency matrices, constructed based on geographical distance or simple river network topology, cannot accurately reflect the implicit connections in actual hydrological processes, thus limiting the accuracy of water level prediction.
The similarity between sites is calculated using the Geographic Optimal Similarity (GOS) principle, an adjacency matrix is constructed, and a graph convolutional neural network (GCN) is used for temporal state updates. Multi-dimensional environmental similarity is captured through multi-kernel graph convolutional layers and gated recurrent mechanisms to construct a graph structure network suitable for river network topology.
It significantly improves the accuracy and robustness of water level prediction, can better quantify the multidimensional environmental similarity between stations, overcomes the problems of single spatial feature extraction and loose spatiotemporal coupling in traditional models, and meets the prediction needs under real-time and complex conditions.
Smart Images

Figure CN122364831A_ABST