The preliminary experiment of genetic-neural network flood forecasting based on QPE and QPF

Taking the Shuibuya control watershed in the upstream of Qingjiang in Hubei Province as an example, in this study we first use grouped Z-I relationships and radar precipitation estimates calibrated by data from surface meteorological stations to calculate the area averaged precipitation of the watershed. Then, genetic algorithms and neural networks method are combined to establish a revised AREM precipitationforecasting model in order to improve forecast accuracy of AREM precipitation. Finally, AREM precipitation data before and after applyingthe revised model are inputted to the Xinanjiang hydrological model to examine the accuracy of the flood forecasts. Results show that the revised AREM precipitation forecasting model can significantly improve the forecast accuracy of the event cumulative precipitation. Theaveraged relative error reduction rate is more than 60%. Hourly precipitation forecast accuracy is also improved to some extent, although there is still some bias compared to actual observations. The averaged flood forecast deterministic coefficient of the AREM precipitation forecast by using the revised model is improved from -32.6% to 64.38%, peak relative error is decreased from 39% to 25.04%. The improvement to thedeterministic coefficient is better than that to the peak relative error. The overall flood forecast accuracy has generally improved.