Data-Driven and Knowledge-Guided Heterogeneous Graphs and Temporal Convolution Networks for Flood Forecasting

Abstract

Data-driven models have been successfully applied to flood prediction. However, the nonlinearity and uncertainty of the prediction process and the possible noise or outliers in the data set will lead to incorrect results. In addition, data-driven models are only trained from available datasets and do not involve scientific principles or laws during the model training process, which may lead to predictions that do not conform to physical laws. To this end, we propose a flood prediction method based on data-driven and knowledge-guided heterogeneous graphs and temporal convolutional networks (DK-HTAN). In the data preprocessing stage, a low-rank approximate decomposition algorithm based on a time tensor was designed to interpolate the input data. Adding an attention mechanism to the heterogeneous graph module is beneficial for introducing prior knowledge. A self-attention mechanism with temporal convolutional network was introduced to dynamically calculate spatiotemporal correlation characteristics of flood data. Finally, we propose physical mechanism constraints for flood processes, adjusted and optimized data-driven models, corrected predictions that did not conform to physical mechanisms, and quantified the uncertainty of predictions. The experimental results on the Qijiang River Basin dataset show that the model has good predictive performance in terms of interval prediction index (PI), RMSE, and MAPE.