Research On Neural Network Forecast Model Of Flood Discharge In Mountainous Watershed

Flood disaster occurs frequently in the middle or small catchment of mountainous watershed with huge losses due to the specific natural,geography as well as the climate.Flood forecast,as a significant part on flood control and hazard mitigation of the middle or small catchment of mountainous watershed,has been a crucial subject for reducing the losses of flood disaster.The neural network is applied to mountainous watershed for developing a neural network model about flood forecasting,and Chongyang river catchment is taken as a case study in this thesis.The main contents are as follows:(1)The basic principles of neural network algorithm,gradient descent back-propagation(gradient descent BP),Levenberg-Marquardt back-propagation(LMBP)and radial basis functions(RBF)are put forward respectively.(2)Then the flood characteristics of Chongyang river catchment are analyzed and flood frequency calculation is carried out at some stations along the river basin.Through a comprehensive analysis on the measured flood information,the flood transmission time from Wuyishan hydrological station in the tributary upstream Chongyang river to Jianyang hydrological station in the lower Chongyang river(?1)and the flood travel time from Masha hydrological station in the tributary of Mayang river to Jianyang hydrological station in the lower Chongyang river(?2)are be determined.(3)Three neural network flood forecasting models are proposed by using gradient descent BP algorithm,LMBP algorithm and RBF algorithm respectively.The three algorithms are then used to establish flood forecast models,selecting the flood discharge of Jianyang hydrological station in the lower Chongyang river and Wuyishan hydrological station in the tributary upstream Chongyang river along with Masha hydrological station in the tributary of Mayang river.Q1(t-?1)and Q2(t-?2)are input to the model and Q(t)is output to the model,where Q1(t-?1)is available as the flood discharge variable of Wuyishan hydrological station in the tributary upstream Chongyang river at ?1 hours before,Q2(t-?2)is available as the flood discharge variable of Masha hydrological station at ?2 hours before and Q(t)is the corresponding flood discharge variable of Jianyang hydrological station in the lower Chongyang river.The relevant parameters about network Parameters are obtained by trial and error.(4)In taking the actual conditions of the basin and prolong the prediction period perspective,an appropriate variable T is determined according to ?1 and ?2(T?>?1 or T??2,generally).Then Q(t-T),which means the flood discharge variable of Jianyang hydrological station in the lower Chongyang river at Thours before,is added to the input of the models.Three neural network flood forecasting models are proposed by using the above mentioned algorithms.Similarly,the relevant parameters about network Parameters are obtained by trial and error.(5)The results indicate that the neural network flood forecast models can provide timely and accurate predictions and thus can be one of the bases for flood forecasting in the flood control departments.It contributes to flood prevention departments with tools to minimize flood disaster in the middle or small catchment of mountainous watershed.Relatively,gradient descent BP algorithm and RBF algorithm are preferable.And the comparisons between models established with different input parameters show that the models considering flood discharge variable of Jianyang hydrological station in the lower Chongyang river at Thours before perform better.The model using gradient descent BP algorithm and considering flood discharge at Thours before performs best.