Practical Data-Driven Flood Forecasting Based on Dynamical Systems Theory

Data-driven flood forecasting methods are useful, especially for rivers that lack information required for building physical models. Although the former methods can forecast river stages using only past water levels and rainfall data, they cannot easily predict unprecedented water levels and require a large amount of data to build accurate models. We focus on phase-space reconstruction approaches based on dynamical systems theory and develop a practical data-driven forecasting method to overcome existing problems. The proposed method can predict unprecedented water levels owing to a proposed correction term, and provide forecasts using only a small number of water level increase events. We applied the proposed method to data from actual rivers and it achieved the best forecast performance among existing data-driven methods, including a multilayer perceptron, and a conventional method based on phase-space reconstruction. In addition, the proposed method forecasted the exceedance of the evacuation warning level 6 h earlier for small and steep rivers. Given its performance and maintainability, the proposed method can be applied to many gauged rivers to facilitate early evacuation.

Automated summary

The study focuses on data-driven flood forecasting methods, particularly for rivers lacking information for physical models. By utilizing phase-space reconstruction approaches based on dynamical systems theory, the proposed method can accurately predict unprecedented water levels with limited data, outperforming existing methods in forecast performance. Additionally, it allows for early evacuation warnings for small and steep rivers, showcasing both its effectiveness and applicability for various gauged rivers.