Fully distributed rainfall-runoff modeling using spatial-temporal graph neural network

Abstract

Recent studies using latest deep learning algorithms such as LSTM (Long Short-Term Memory) have shown great promise in time-series modeling. There are many studies focusing on the watershed-scale rainfall-runoff modeling or streamflow forecasting, often considering a single watershed with limited generalization capabilities. To improve the model performance, several studies explored an integrated approach by decomposing a large watershed into multiple sub-watersheds with semi-distributed structure. In this study, we propose an innovative physics-informed fully-distributed rainfall-runoff model, NRM-Graph (Neural Runoff Model-Graph), using Graph Neural Networks (GNN) to make full use of spatial information including the flow direction and geographic data. Specifically, we applied a time-series model on each grid cell for its runoff production. The output of each grid cell is then aggregated by a GNN as the final runoff at the watershed outlet. The case study shows that our GNN based model successfully represents the spatial information in predictions. NRM-Graph network has shown less over-fitting and a significant improvement on the model performance compared to the baselines with spatial information. Our research further confirms the importance of spatially distributed hydrological information in rainfall-runoff modeling using deep learning, and we encourage researchers to incorporate more domain knowledge in modeling.

Real-Time Streamflow Forecasting Framework, Implementation and Post-Analysis Using Deep Learning

“…On the other hand, our framework and models were developed for forecasting at the watershed level, and we neglected the geographical distribution of rainfall. Several simulation experiments involving graph neural networks have been conducted (Jia et al, 2021;Xiang et al, 2022), and that could be addressed in future studies. We recommended updating the deep learning model continuously with the most recent data in our framework as a potential improvement step for prediction.…”

Spatial Downscaling of Streamflow Data with Attention Based Spatio-Temporal Graph Convolutional Networks

“…Utilization of graph neural networks in hydrological context has been appealing to domain scientists lately. While Sit et al [36] forecasted hourly stream ow for a basin in Iowa using upstream information, Xiang and Demir [37] showed that a similar approach could be used to generalize forecasts for the whole state. In a similar approach, Jia et al [38] showed that recurrent graph models outperformed LSTM networks for both water ow and temperature in a transfer learning approach.…”

Enhancing Hydrological Modeling with Transformers: A Case Study for 24-Hour Streamflow Prediction

“…However, advancements in artificial intelligence (AI) coupled with the increasing capabilities of graphics processing units (GPUs) have opened up new possibilities and accelerated the progress of deep learning techniques, which has led to the widespread usage of these techniques in streamflow forecasting as well (Sit et al, 2022a). Out of various neural network architectures explored for streamflow forecasting (Sit et al, 2021a;Xiang and Demir, 2022b;Chen et al, 2023), Recurrent Neural Networks (RNNs), especially the Long Short-Term Memory (LSTM) neural network and Gated Recurrent Units (GRUs), have emerged as the most extensively studied and researched models in this domain. Kratzert et al (2018) applied an LSTM model to predict daily runoff, incorporating meteorological observations, and demonstrated that the LSTM model outperformed a wellestablished physical model in their study area.…”