Research on flood forecasting method in mountainous small watersheds based on machine learning for identifying rainfall dynamic spatiotemporal features

The mountainous region experiences fast — flowing and highly destructive floods, posing challenges for accurate and timely forecasting. Enhancing the accuracy and lead time of flood prediction in mountainous areas is a pressing issue. Addressing this concern, this paper proposes an innovative flood forecasting method based on machine learning technology. The approach identifies historical rainfall-flood events with the most similarity to the current dynamic spatiotemporal features of rainfall, employing a learn from the past to predict the present strate-gy. The results indicate that, in small watersheds with minimal human influence and a basin area of approximately 600 km in mountainous regions, the method not only predicts the overall trend of rainfall but also forecasts the asso-ciated mountainous flood processes under this rainfall trend. The average errors for peak flow, flood volume, and peak time are 8.33%, 14.27%, and 1 hour, respectively, meeting the accuracy requirements for flood forecasting. Distinguished from traditional flood forecasting methods, this approach predicts mountainous floods from the perspective of the overall rainfall trend, providing a targeted strategy for flood forecasting in small watersheds in hilly areas. © 2024 International Research and Training Center on Erosion and Sedimentation and China Water and Power Press. All rights reserved.