HUANG Yan-chun ,CUI Ning-bo,CHEN Xuan-quan,XU Hao-ruo,ZHANG Yi-xuan
China Rural Water and Hydropower. 2020, (5):
13-20.
Knowledge map
 
Save
In order to explore the adaptability of different numerical simulation models in the hilly area of central Sichuan and improve the ET0 prediction accuracy under the condition of missing meteorological data. This study constructs four input combinations of different meteorological factors with the daily meteorological data from 7 representative stations from 1961 to 2016 and establishes numerical simulation models based on M5 regression tree (M5-RT), back propagation neural network optimized with double hidden layers (H-BPNN) and generalizes regression neural network of cross validation (CV-GRNN) optimization. At the same time, three empirical models (Jensen- Haise, Hargreaves-Li and Irmak-Allen models) with higher precision in the hilly area of central Sichuan are selected to compare with the numerical simulation models. Finally, this paper evaluates the ET0 prediction accuracy of different models on a daily scale and evaluates the generalization ability of the three types of numerical simulation models in the hilly area of central Sichuan with portability analysis. The results are as follows: M5-RT2, CV-GRNN2 and H-BPNN2 models based on temperature, wind speed and extraterrestrial radiation have high precision, with their determination coefficients (R2) reaching 0.987, 0.967 and 0.988 respectively, and Nash-Sutcliffe efficiency coefficient (NSE) reaching 0.987 1, 0.937 4 and 0.988 8 respectively. The three models can fit the complicated nonlinear mapping relationship between ET0 and meteorological parameters to a certain extent. The daily scale error analysis shows that M5-RT regression tree model is the best, H-BPNN model is the second, and CV-GRNN model is the worst, but the root-mean-square errors (RMSE) are less than 0.5 mm/d and the average relative errors (MRE) are less than 13.59% in all three models. The accuracy of the numerical simulation models is higher than that of the empirical models, among which, the M5-RT model input into combination 2 (extraterrestrial radiation, maximum/minimum temperature and wind speed), combination 3 (maximum/ minimum temperature and wind speed) and combination 4 (extraterrestrial radiation and wind speed) has wide adaptability in the simulated forecast of ET0 with missing meteorological data. Portability analysis shows that the prediction accuracy of the numerical simulation models have decreased under the cross combination of training and prediction stations, among which, M5-RT model has the strongest generalization ability, its simulation output has the highest stability. The H-BPNN model and CV-GRNN model have a large truncation error when ET0 is greater than 6 mm/d, the predicted value is generally small. At the same time, the slope of the trend line of the simulation result and the standard value of the CV-GRNN model is small, and the simulation value is small overall. The ET0 prediction models based on M5 regression tree have high accuracy and stable simulation results, which can be recommended as a simplified model for predicting ET0 in the hilly area of central Sichuan.