Riverine Flood Forecasting Using Advanced Deep Learning Approaches
TL;DR
Analysis of riverine flood forecasting models revealed that PatchTST outperformed the other models during moderate‐flow regimes while falling behind during extreme flooding events, and sensitivity analysis results indicated that PatchTST was slightly more sensitive to the selected training data features.
Abstract
Accurate riverine flood forecasting is crucial for effective river management. This paper utilized the Time‐Series Dense Encoder (TiDE), Neural Hierarchical Interpolation for Time Series Forecasting (N‐HiTS), and Patch Time Series Transformer (PatchTST) to forecast riverine flood and benchmarked their results against Long Short‐Term Memory (LSTM). Each model was implemented with varying forecast lead times for the Proctor Creek–Chattahoochee watershed, Georgia, USA. Additionally, the sensitivity of each model was evaluated by excluding meteorological forcing features one by one to determine how the performance varied across different variables. The time of concentration () was incorporated as a physical parameter in the algorithm's lookback window. The trained models were evaluated separately on event‐based simulations. The Diebold–Mariano statistical test was utilized for a thorough analysis of performance. Analysis revealed that PatchTST outperformed the other models during moderate‐flow regimes while falling behind during extreme flooding events. An 18‐ to 24‐h lookback window was found to be statistically optimal for the models. The sensitivity analysis results indicated that PatchTST was slightly more sensitive to the selected training data features. Incorporating into the lookback window revealed that TiDE and LSTM showed better performance when using a sequence length of 18 h while PatchTST achieved the best results with a 24‐h lookback window.