Short-term prediction of reference crop evapotranspiration based on machine learning with different decomposition methods in arid areas of China

Abstract

This paper established hybrid prediction models based on variational mode decomposition (VMD), empirical mode decomposition (EMD), and ensemble empirical mode decomposition (EEMD) combined with the backpropagation neural network model (BPNN) to improve the prediction accuracy of reference crop evapotranspiration (ET0) time series with the characteristics of nonlinearity and instability. The daily ET0 data of 11 representative stations in Xinjiang from 1993 to 2016 were selected for model training and testing and compared with the results of support vector regression (SVR) and gradient boosting regression tree (GBRT) as the two empirical machine learning models. The results indicated the superiority of the VMD-BPNN hybrid prediction model to EMD-BPNN and EEMD-BPNN in terms of accuracy and stability, with root mean square error (RMSE) = 0.405 mm/d, mean absolute error (MAE) = 0.268 mm/d, and coefficient of determination (R2) = 0.979. When employing the VMD-BPNN model to forecast ET0 for seven days, the RMSE and Nash-Sutcliffe efficiency coefficient (NSE) of the VMD-BPNN model were 0.588 mm/d and 0.952, respectively, and the prediction results indicated high precision and reliability. The prediction accuracy of the VMD-BPNN model was significantly higher than that of single machine learning models, such as BPNN, SVR, and GBRT. The RMSE and MAE values of the VMD-BPNN model were more than 60% smaller than BPNN, SVR, and GBRT models, and R2 and NSE were approximately 18% higher than BPNN, SVR, and GBRT models, respectively. This demonstrates the effectiveness of the VMD method in reducing the non-stationarity of the original daily ET0 data. The BPNN model predicted the decomposed data series, and the prediction accuracy and stability were significantly enhanced. This indicates the high reliability of the VMD-BPNN model and its capability for ET0 prediction in Xinjiang.