Abstract: Real-time correction, as the last barrier to improving forecast accuracy, is an important part of flood forecasting. To address the problem of poor correction of flood process and flood elements, a joint real-time correction method combining a Kalman filter and a K-nearest neighbor algorithm is developed through selecting the rainfall runoff series of the basin and the eigenvalues of its historical flood data and constructing a storm flood feature database, in combination with a case study of the watershed above Tunxi of Hengjiang River. The results show that relative to the uncorrected model or single correction method, this new method is more effective in reducing forecast errors of flood peak, volume, and flood peak time, while it can maintain stability and accuracy for a forecast period of six hours or shorter. It could play an important role in improving flood forecast accuracy, effective early warning, and disaster prevention and mitigation for small- and medium-sized rivers or similar regional flood forecasting.

Key words: flood forecast, real-time correction, K-nearest neighbor algorithm, ensemble Kalman filter, storm and flood database, Tunxi River basin