Study on Ensemble Calibration of Flood Forecasting Based on Response curve of rainfall dynamic system and LSTM.

Abstract

Abstract A dynamic system response curve correction method is used to establish the error of the rainfall time series, modify the surface precipitation of the basin, and use the surface precipitation of the revised as input to improve flood forecast accuracy using the Shouxi watershed in Sichuan Province and the Qingyangcha domain in Shanxi Province as the research object. Flood forecasting is carried out using the excess storage and excess infiltration simultaneously model. Combined with the hierarchical optimization method and the LSTM (Long Short-Term Memory neural network) error output correction method, the three-process error set correction is carried out. The results show that the accuracy of the flood forecast discharge with area-rainfall correction was improved compared with that without area-rainfall correction. Specifically, The absolute value of flood peak error for 12 of 15 floods in the Shouxi Basin validation set decreased by 0.566.3% compared to before the areal rainfall correction, the NSE of 13 flood forecast discharge increased by 0.0020.015, and the current difference of 2 peaks shortened by 1 hour. The peak error of five floods was reduced by 0.23-5.49% in the Qingyangcha watershed validation set compared to before the areal rainfall correction, the NSE (Nash Social Welfare Efficiency) of five flood forecast flows was increased by 0.01-0.071, and the current difference of two peaks was shortened by one hour. The comprehensive results show that this method is applicable to reduce the prediction error and improve the accuracy of flood forecasting in the watershed.