Monthly runoff prediction by a multivariate hybrid model based on decomposition-normality and Lasso regression.

ABSTRACT

The intensified non-stationary, skewness, non-linear nature of runoff series due to the comprehensive influences of meteorological events and human activities has brought new challenges to accurate runoff prediction. To solve the issues, a multivariate hybrid model introducing decomposition-normality mode into SVR was proposed. The normal transformation techniques, Box-Cox transformation, and W-H inverse transformation were employed to transform the input variables of the model into normal distribution to overcome the error caused by skewness of the runoff data. The results show that decomposition-normality mode can improve the performance of the models. In particular, WT-BC-LSVR accurately predicted peak flow and low flow during the testing, and the mean relative errors are less than 16%, Rs and Nash-Sutcliffe efficiencies are greater than 0.97 and 0.94, respectively. The study demonstrates that the proposed multivariate hybrid model based on the decomposition-normality mode is a novel promising prediction model with satisfactory performance that can accurately predict complex monthly runoff.