Research On Daily Runoff Forecast Based On Cascaded Long And Short-term Memory Mode

Accurate streamflow forecast plays an important role in water resources planning and management,flood forecast,and early warning.Affected by many factors such as precipitation,soil moisture,and evapotranspiration,the streamflow change is highly nonlinear,leading to a grand challenge in streamflow forecasting especially at long lead times.Currently,machine learning techniques have been widely applied to streamflow prediction at short lead times,and increasing numbers of predictors are introduced to improve predictive skills.However,the effect of different predictors and their combinations on the prediction at long leads remains to be explored.In addition,medium and long-range streamflow forecasts largely depend on the accuracy of meteorological forecasts.Due to large errors in precipitation forecasts,most streamflow forecasts based on deep learning rely only on historical data.Therefore,the research objective of the paper is to carry out daily streamflow forecast at the mainstream of the Yangtze River based on the Long Short-Term Memory(LSTM)model,taking into account different forecasting factors and their combinations.Further,a cascade LSTM model is established to first predict precipitation,in order to explore the prediction performance and potential of the model’s medium and long-term daily synthetic streamflow.Focusing on the above research objectives,this paper has carried out work on the daily streamflow prediction by using machine learning method,and the main conclusions are as follows:(1)Different combinations of forecast factors have different effects on streamflow forecast at different lead times.Precipitation can increase the accuracy of streamflow forecasting,while soil moisture and evapotranspiration can further improve the accuracy of streamflow forecasting during long lead times.The LSTM model is used to forecast the daily streamflow of four hydrological stations(Zhimenda,Cuntan,Hankou,Datong)over the Yangtze River during flood seasons at lead times of 1～30 days,by using different combinations of historical streamflow,precipitation,soil moisture,and evapotranspiration as predictors.The effects of different predictors on daily streamflow and high/low flows,together with the difference between short and long lead times,are investigated.The reference experiment which only uses streamflow predictor has a high(0.58～0.99)Kling-Gupta efficiency(KGE)at 1～7days lead times,but its performance decreases rapidly as the lead time increases.Adding precipitation predictor increases the KGE of daily streamflow by 0.09～0.21 and corrects negative/positive biases in high/low flows,with stronger positive effects occurring at longer lead times(>20 days).Soil moisture and evapotranspiration predictors have added values to daily streamflow forecasting at long lead times by further improving the KGE by 0.04～0.11,but exert minor influences on streamflow extremes.(2)We apply a cascade Long Short-Term Memory(LSTM)model to forecast daily streamflow over 49 control stations in the Yangtze River basin,it is found that compared to the LSTM model,the cascade LSTM model show higher accuracy in streamflow forecasting in more than 30 control stations of the Yangtze River..The first layer of the cascade LSTM model uses atmospheric circulation factors to predict future precipitation,and the second layer uses forecast precipitation to predict streamflow.The land surface hydrological model is used to generate continuous streamflow series at 49 control stations for training the LSTM model.It is found that the LSTM model has high streamflow forecast skills at most control stations.At the lead times of 1,7,and 15 days,the streamflow Kling-Gupta efficiency(KGE)values are greater than 0.5 over 78%,30%,and 20% watersheds,respectively.Its performance improves with the increase in drainage area.After implementing the cascade LSTM model,61%～88% of the watersheds show increased KGE at different leads,and the increase is more obvious at longer leads.Using cascade LSTM with perfect future precipitation shows further improvement,especially over small watersheds.It shows that the cascade LSTM model has potential in the streamflow forecast over the Yangtze River basin.