Application of SVR Model Based on Box-Cox Transformation and Lasso Regression in Monthly Runoff Prediction

In order to improve the accuracy of the medium-long term runoff prediction and provide more reliable information support for reservoir operation decision and water resource allocation management, the Box-Cox transform and Lasso regression methods were introduced into support vector regression (SVR) model according to the skewness and nonlinearity characteristics of runoff series. A new hybrid model based on Box-Cox transform and Lasso regression and support vector regression (SVR) called BC-LSVR was proposed. On the basis of the phase space reconstruction of the original runoff series and Box-Cox normalization transformation, Lasso regression was used to identify the forecast factors, then predicted monthly runoff. The MM-LSVR and BC-GSVR prediction models were constructed by using the Min-Max data standardization method and the Gray forecast factor identify method as comparison methods, respectively, and the comparative analysis of the prediction effect of three models was carried out. The models was applied to six main hydrological stations such as Linjiacun in Weihe basin, etc. The results show that the BC-LSVR model has the best prediction effect. In the validation period, the average absolute relative error (MARE) of the six stations is less than 20%, the qualification rate (QR) is greater than 0.6, and the efficiency coefficient (Ens) is greater than 0.52. In the calibration period and validation period, three evaluation indexes of the BC-LSVR model are superior to those of the MM-LSVR and BC-GSVR models, indicating that the introduction of Box-Cox transformation and Lasso regression can effectively improve the prediction accuracy of the SVR model. This paper is expected to provide an effective method to improve the accuracy of medium-long-term runoff prediction. © 2022, The Editorial Board of Journal of Basic Science and Engineering. All right reserved.