Error-correction-based data-driven models for multiple-hour-ahead river stage predictions: A case study of the upstream region of the Cho-Shui River, Taiwan

Abstract

Study Region: The Cho-Shui River Basin is located in central Taiwan and has a catchment area of approximately 3157 km2 and an annual average rainfall of 2200 mm. It is the longest river in Taiwan, and its mainstream has an average slope of 0.018. However, the prediction of flash floods upstream of the Cho-Shui River has yet to be investigated. Because of the river’s steep slope, typhoons or storms can result in significant flood disasters. Therefore, accurate and reliable river stage prediction is required for flood disaster mitigation. Study Focus: A predictor-corrector multiple-hour-ahead methodology was developed and used for constructing seven data-driven models for river stage predictions. The proposed methodology employs both river-stage and residual-error prediction models. The optimal residual-error prediction model was determined using seven data-driven models. An extensive comparison of the seven data-driven models was conducted regarding prediction performance with 1–24-h lead times. New Hydrological Insights for the Region: The proposed error-correction-based data-driven models exhibited high prediction accuracy, making them suitable for improvements of river stage predictions with the decrease in lead-time-averaged root mean square error (RMSE), up to − 73.7 %. The error-correction-based categorical gradient boosting regression (CGBR) and encoder-decoder long short-term memory (LSTM) models outperformed the other models, yielding the average peak water-level error (PWE) of 0.3 m for the 24-hour-ahead prediction. Thus, these two models could be helpful for early-warning river flood forecasting during typhoons or storms.

Keywords

• Multiple-hour-ahead
• River stage
• Machine learning
• Ensemble learning
• Deep learning