Self-organizing maps of typhoon tracks allow for flood forecasts up to two days in advance

Taiwan's mountainous terrain and location in the path of typhoons make it highly susceptible to devastating floods. The short, steep rivers result in rapid flow increases within hours of typhoon passage, overwhelming small reservoirs and causing significant flood hazards. Existing flood warning systems provide only a few hours' notice, insufficient for effective reservoir management and flood mitigation. The increasing frequency and intensity of typhoons due to climate change further exacerbate this problem. Accurate, advanced flood forecasting is crucial for managing reservoir capacity for both flood control and water supply. While current rainfall-runoff models provide short-term (6 hours) forecasts, a longer lead time is required for effective reservoir operation. This study aims to develop an AI-based method to improve flood forecasting lead times to at least two days, using historical typhoon track data and real-time remote sensing information.

Previous research has employed K-means and fuzzy clustering to group typhoon tracks based on their paths over Taiwan. However, these methods lack the spatial resolution and digital representation necessary for detailed analysis of track-terrain interactions and their influence on rainfall intensity. Recent studies have shown success in using AI techniques for site-specific rainfall-runoff prediction for individual typhoons. These AI-based approaches leverage massive historical datasets and real-time remote sensing data, but they often focus on short-term forecasting. This study addresses the need for improved long-term flood prediction by integrating typhoon track characteristics and flow data.

This research utilizes a four-module AI-based methodology to predict flood hydrographs for the Shihmen Reservoir watershed in Taiwan. The methodology begins with **typhoon track vectorization**, converting analog typhoon tracks into digital vectors using a grid system that accounts for topographic variability. A diffusion process is applied to the track vectors to incorporate the spatial relationships of typhoon impacts. Next, **track vector clustering** is performed using a self-organizing map (SOM) to group similar typhoon tracks into clusters. The SOM algorithm groups tracks with similar shapes and paths, reflecting similar rainfall characteristics and reservoir inflows. The third module, **FCC extraction**, involves creating flow characteristic curves (FCCs) for each cluster. FCCs are cumulative curves showing the percentage of time specified discharges were equaled or exceeded during a typhoon event. These are normalized to account for differences in total flow volume and typhoon duration. Three different schemes for defining typhoon duration are explored to optimize the similarity of FCCs within each cluster. These schemes consider the arrival and departure times of the typhoon, the start of significant flow increase and typhoon departure, and the start of significant flow increase and rainfall cessation. The final module, **flood hydrograph prediction**, uses the forecasted track of an approaching typhoon to identify the best-matched cluster in the SOM. Flood hydrograph prediction is then achieved using either the FCC of the best-matched typhoon track or the average FCC of all tracks within that cluster and the predicted total rainfall. Model training uses data from 87 typhoons between 1965 and 2015, and the model's reliability is tested using data from 10 events in 2013 and 2019.

The study found that the SOM-based approach accurately predicts flood hydrographs up to two days before typhoon landfall. The use of the 4x4 SOM effectively clusters typhoon tracks, revealing distinct relationships between track patterns and resulting reservoir inflows. The three different schemes for defining typhoon duration are investigated, with the third scheme (duration between significant flow increase and rainfall cessation) yielding the most consistent and accurate FCCs. Both strategies for selecting FCCs (best-matched track and average cluster FCC) demonstrate good predictive power. The model demonstrates significant improvement over traditional rainfall-runoff models, which typically provide only short-term forecasts. Testing against 10 independent typhoon events showed that the model accurately predicts both the timing and volume of peak flows. Comparison with a storage function model (SFM) further confirms the superiority of the SOM-based approach, particularly for events with high peak flows. Even with typhoon track prediction errors up to 80 km, the model demonstrates acceptable accuracy, highlighting its robustness and error-tolerance. The study also showed that the strategy based on the best-matched typhoon track generally yielded more favorable results.

The results demonstrate the effectiveness of integrating typhoon track information with a machine-learning approach for long-lead-time flood forecasting. The ability to predict flood hydrographs up to two days in advance offers significant improvements to existing warning systems, providing valuable time for reservoir management, flood defense implementation, and public safety measures. The superior performance compared to traditional rainfall-runoff models highlights the advantages of this AI-based approach in capturing the complex spatiotemporal variability of typhoon-induced rainfall. The model's robustness to errors in typhoon track prediction makes it a practical tool for real-world applications. The approach developed in this study serves as an effective decision support tool for water resource management, bridging the gap between typhoon track prediction and flood forecasting.

This study presents a novel AI-based methodology that significantly enhances flood forecasting capabilities in typhoon-prone regions. By integrating typhoon track data with a self-organizing map and flow characteristic curves, the model successfully predicts flood hydrographs up to two days in advance. This surpasses the capabilities of traditional methods and provides a valuable tool for reservoir management and flood mitigation. Future research could focus on incorporating additional variables (e.g., rainfall intensity, wind speed) to further refine prediction accuracy, and extending the lead time by improving the accuracy of typhoon track predictions.

The model's accuracy depends on the accuracy of the typhoon track forecast. While the model shows some tolerance to prediction errors, significant errors could impact the accuracy of the flood hydrograph prediction. The model was developed and tested for a specific watershed; its generalizability to other watersheds needs further investigation. The selection of optimal weights in the diffusion process was based on trial-and-error and may benefit from further optimization.