Comparison of three updating models for real time forecasting: a case study of flood forecasting at the middle reaches of the Huai River in East China

Abstract

This study explores the performances of three real-time updating models in improving flood forecasting accuracy. The first model is the K-nearest neighbor (KNN) algorithm. The KNN algorithm estimates forecast errors based upon the most similar samples of errors rather than on the most recent ones. The two other updating models are the Kalman filter (KF) and a combined model incorporating both the KF and KNN procedures. To compare the performances of these three models, this study uses the middle reaches of the Huai River in East China for a case study. Using 13 flood events occurring from 2003 to 2010 as examples, one hydraulic routing model is applied for flood simulation. Subsequently, the three updating models are utilized with lead times of 1- to 8-h for updating the outputs of the hydraulic model. Comparison of the updated results from the three updating models reveals that all three updating models improve the performance of the hydraulic model for flood forecasting. Among them, the KNN model performs more robustly for forecasts with a longer lead time than the other two updating models. Statistical results show that the KNN model is capable of providing excellent forecasts with an 8-h lead time in both the calibration and validation periods.