Abstract
Three types of watershed spatial attributes are separately used to guide dynamic edge-weighted GCN (DynE-GCN) in learning spatial dependencies between each station's observed rainfall. Temporal dynamics are modeled by LSTM, followed by a VAE that compresses high-dimensional spatiotemporal features and reconstructs new data. Finally, by modulating the VAE output and sampling frequency, both deterministic and probabilistic flood forecasting can be performed. AI-based forecasting models have emerged as a hot topic in flood forecasting. To address the limitations of single models in uncertainty quantification and flood peak prediction, this study proposes a novel spatial-temporal AI model, GCN-VAE-ResLSTM. The model improves LSTM by utilizing a Graph Convolutional Network (GCN) with dynamic edge weights, termed DynE-GCN, to learn spatial dependencies of rainfall. Three residual connections are employed to strengthen information flow. A Variational Autoencoder (VAE) is incorporated for both extreme trend learning and uncertainty modeling. Seven series models are constructed to investigate the contributions of all modules to LSTM. Taking the Shifeng River basin in Zhejiang Province, China, as the study area, this study incorporates three watershed spatial attributes, namely hydrological network distribution, topography, and distance. Results indicate: (1) The DynE-GCN exhibits superior adaptability to spatial feature inputs. (2) The GCN-VAE-ResLSTM delivers optimal performance for deterministic predictions, reducing RMSE and PRE by 56.4 and 60.7%, respectively, compared to LSTM at a 1-hour lead time. And the GCN-VAE-LSTM delivers robust and accurate flood peak predictions at all lead times. (3) In probabilistic forecasting, the GCN-VAE-ResLSTM demonstrates the highest accuracy, while the VAE-ResLSTM achieves better sharpness, reliability, and timeliness.
Funding Information
  • the Zhejiang Key Research and Development Program (2025C02049)

This publication has 58 references indexed in Scilit: