Models Of Runoff Forecast In Jinghe River Basin

Runoff is affected by many factors such as climate, geographical and human activities, its characteristic is very complicated, showing randomicity, grey, nonlinear, and so on. Accurate and timely runoff prediction for the reasonable allocation and make full use of water resources as well as flood control and drought relief has important theoretical significance and guiding value. Irrigated agriculture of Jinghe River is developed, but the uneven distribution of surface runoff during the year cause agricultural irrigation in the critical period of water shortage. The runoff forecast can provide a basis of making rational use of water resources in river basin and irrigation real-time scheduling.Based on the streamflow data from Zhangjiashan Station on Jinghe basin, prediction of monthly streamflow and daily streamflow is developed. In the prediction of monthly streamflow part, four single forecasting models were introduced: BP neural network model, support vector machine regression model(SVM), generalized regression neural network model(GRNN) and auto regressive integrated moving average model(ARIMA). In combination forecasting model, Combined with multi-resolution analysis of wavelet analysis theory, three kinds of combination forecast method is established: wavelet neural network combined model(WANN), wavelet support vector regression combined model(WSVM), wavelet generalized regression neural network combined model(WGRNN). Verification and calibration of models were realized by MATLAB software. The root mean square error(RMSE), mean absolute error(MAE), deterministic coefficient(DC) and correlation coefficient(R) statistics are used for the comparing criteria of accuracy. Comparing the accuracy of various prediction models, the optimal model which is preferably suitable for Zhangjiashan station in monthly runoff forecast is selected.And then using the optimal combination method to predict the daily runoff.Results are as follows:(1) Comparing the three single models, BP model、SVM model and GRNN model,which their prediction time is the same length, we can find that the SVM model with accuracy of RMSE=27.93m3/s, MAE=13.43m3/s, DC=0.338,R=0.662 in test period,which is superior than the accuracy of the other two models.Thus when the forecast period is longer, we can give priority to SVM model. ARIMA model takes into account the effects of different years of the same month, can be applied shorter forecast.(2) In order to improve the accuracy of the individual prediction model, this paper attempts to combine wavelet analysis with the first three single forecasting methods. Taking full advantage of multi-resolution analysis principle of wavelet, WANN model, WSVM model and WGRNN model are established. Though the results we found that the combination model can greatly improve the accuracy of single forecast model. The four evaluation indexes of WANN model in test period were 26.03m3/s, 17.96m3/s, 0.425, 0.664, better than the indexes of BP neural network model in the test period. And accordingly, the four evaluation indexes of WSVM model were12.46 m3/s, 7.74m3/s, 0.868, 0.935, respectively, which are far superior to SVM model. The four indexes of WGRNN model are 18.91m3/s, 13.14m3/s, 0.697, 0.854, which are far superior to GRNN model.(3) Comparing of the three combined forecasting model, we found that the accuracy of WSVM and WGRNN model satisfied monthly runoff forecast demanded accuracy and in the testing stage of the indicators is better than WANN model, and the precision of WSVM is highest, which significant that WSVM model is the most optimal forecasting model for monthly runoff prediction of Zhangjiashan Station and it can be applied to predict the monthly runoff of Zhangjiashan station.(4) To verify the validity and practicality of WSVM model, this paper attempts to apply the combined model to predict the daily runoff of Zhangjiashan station. From the results we can see that for WSVR model in test period, the indexes such as RMSE, MAE, DC and R were 26.05m3/s, 8.26m3/s, 0.826 and 0.910, respectively. The accuracy of coupled model was much higher than that of support vector machine model, and it was more accurate in flood season than in dry season. It suggested that the WSVR model could be applied to forecast daily runoff effectively, which provided a new way for runoff forecast.