Real-time Neural-network-based Ensemble Typhoon Flood Forecasting Model with Self-organizing Map Cluster Analysis: A Case Study on the Wu River Basin in Taiwan

Accurate hourly real-time flood forecasting is necessary for early flood warning systems, especially during typhoon periods. Artificial intelligence methods have been increasingly used for real-time flood forecasting. This study developed a real-time flood forecasting model by using back-propagation networks (BPNs) with a self-organizing map (SOM) to create ensemble forecasts. Random weights and biases were set for the BPNs to learn the characteristics of a catchment system. An unsupervised SOM network with a classification function was then used to cluster representative BPN weights and biases; clusters of BPNs with high accuracy were selected to act as experts for the ensemble models to forecast flow rates. The model was applied to flood events in the Wu River Basin of Taiwan. Most observed values were within the forecasting intervals of the BPN clusters in the calibration and validation phases, indicating that the models had acceptable accuracy. For the large flood events of typhoons Saola in the calibration phase and Soulik in the validation phase, the mean average error of the ensemble mean model for the cluster A was 143.1 and 327.4 m(3)/s, respectively; these values were lower than those for the best individual model within the cluster (194.3 and 917.9 m(3)/s). The ensemble model thus outperformed the individual models and can accurately forecast flood values and intervals. Therefore, the model can be used to accurately forecast floods.

Automated summary

This study developed a real-time flood forecasting model using back-propagation networks and self-organizing map for ensemble forecasts. The model was applied in the Wu River Basin of Taiwan and showed acceptable accuracy. The ensemble model outperformed individual models, particularly during typhoon periods, and accurately forecasted flood values and intervals.