Simulation of the flood process in the middle reaches of the Yellow River by a long - short term memory (LSTM) neuro network

Flood forecasting is an important non-engineering measure for flood prevention and disaster reduction. Hydrological situation in the middle reaches of the Yellow River has changed significantly in the past 50 years. To improve accuracy of flood forecasting in semi-arid and semi-humid areas in the middle reaches of the Yellow River, an in-depth learning neuro network model with good temporal resolution was used. The LSTM model was applied to data of 98 storms and floods from 1956 to 2014 at Jingle Control Station, located in the upper reaches of the Fenhe River. Rainfall data from 14 stations and hydrological data at Jingle Station were used as input; flood process under different foreseeable periods was used as output, with a regular rate of 78 events, and verification period of 20 events. Prediction accuracy was found to be higher when foresight period was 0 - 6 h, but relatively poor when foresight period was > 6 h. Prediction accuracy was found to increase with increased number of neurons and trainings. Prediction accuracy increased significantly when forecast period was increased from 0 to 6 h, but increased more uniformly when forecast period was > 6 h.