Spatio-Temporal Evolution Law And Prediction Of Reference Crop Evapotranspiration In Wei River Basin

In recent years,due to the influence of natural factors and human activities,the temporal and spatial distribution characteristics of meteorological elements have changed greatly,which brings great challenges to the accurate prediction of reference crop evapotranspiration（ET₀）.ET₀ is important for regional water resources planning and irrigation scheduling design.The FAO-56 Penman-Monteith model（P-M）recommended by FAO is used as a reference model for predicting ET₀,but its application is generally limited in many parts of the world due to the lack of complete meteorological data.Wei River basin,the largest tributary of the Yellow River,is an important source of water for agriculture,industry and domestic use in Northwest China.To realize efficient utilization and distribution of water resources in this region,it is very important to improve the ET₀ prediction accuracy of Wei River Basin.Taking 10 meteorological stations in the Wei River Basin as research objects,using two data preprocessing technologies,namely Variational Mode Decomposition（VMD）and Boc-Cox transform,four screening methods,such as least absolute shrinkage and selection operator regression（LASSO）,maximum mutual information（MIC）,Random Forest（RF）and recursive feature elimination（RFE）,and five machines,namely extreme learning machine（ELM）,support vector machine（SVM）,BP neural network,Random forest（RF）and short and long time memory network（LSTM）,were used to build multiple hybrid models.The ET₀ prediction model framework based on“decomposition-transformation-recognization-prediction”is proposed.Compared with the performance of two empirical models（Hargreaves-Samani and Priestley-Taylor models）,the following main conclusions are obtained.（1）By means of spatio-temporal law analysis and driver identification,the evolution characteristics of meteorological elements and the main meteorological elements affecting ET₀were studied.The results showed that the annual mean temperature,sunshine duration and wind speed in the Wei River basin showed an upward trend,while the relative humidity（RH）showed a significant downward trend,and ET₀ showed an upward trend of“strong decreasing-strong increasing-weak decreasing-strong increasing”.The highest correlation between meteorological elements and ET₀ is the mean maximum temperature,followed by the mean minimum temperature and sunshine duration.The sensitivity order was RH>SSD>T_max>u>T_min;In the correlation analysis of multiple meteorological elements combination and ET₀,the mutual information value of combination T_max-T_min-RH-SSD and combination T_max-T_min-u-SSD is relatively high,and the MI value of two combinations is about1.7 or more.Therefore,the two combinations can be considered as the input factors for the estimation of ET₀ in the area lacking meteorological data.（2）The traditional single machine learning model is constructed to predict ET₀.The results show that,no matter which combination of input,the prediction accuracy is from high to low:SVM>BP>ELM>RF.The prediction accuracy of SVM-4 model was the highest,with Nash efficiency coefficient（NSE）and correlation coefficient（R）greater than 0.87 and 0.93,respectively,and mean absolute percentage error（MAPE）and root mean square error（RMSE）less than 19.66%and 0.94mm/d,respectively.SVM has the best prediction effect among the traditional single model,indicating that the SVM model has good robustness and strong generalization ability.（3）A single data processing（VMD and BC）coupled with traditional machine learning model was constructed to predict ET₀.The prediction effect evaluation showed that both the mixed model based on VMD decomposition and the hybrid model based on BC transformation had better prediction effect than the single model.Comparing the two data processing techniques,the prediction model with BC transformation has the highest accuracy,among which BC-SVM-4 has the best prediction performance in all models,with NSE and R greater than 0.96 and 0.98,and MAPE and RMSE less than 12.22%and 0.51mm/d,respectively.（4）By integrating VMD,BC and LASSO data processing with machine learning models,a hybrid model based on"decomposition-transformation-recognization-prediction"was constructed to improve the prediction accuracy of ET₀.The results showed that compared with single processing,the hybrid data preprocessing based on VMD-BC was more effective in improving the model performance.In the traditional machine learning model,VMD-BC-SVM-4 model has higher accuracy,where NSE and R are greater than 0.97 and 0.95,MAPE and RMSE are less than 13.0%and 0.45mm/d,respectively.Compared with H-S and P-T empirical models,VMD-BC-SVM model has higher simulation accuracy and can be used as an alternative to ET₀ prediction in regions with scarce meteorological data.LSTM deep learning method has higher accuracy than traditional machine learning model,among which the MAPE and RMSE of VMD-BC-LSTM-4 model are below 9.97%and 0.45mm/d,respectively.However,compared with LSTM,VMD-BC-SVM model not only achieves relatively high accuracy,but also has short running time and is a cost-effective model.