Application of variable selection in hydrological forecasting based on Partial Least Squares

In this paper, a variable selection method based on Partial Least Squares is discussed and used to establish the hydrological forecasting model. The Q2cum score is taken as objective function, while Partial Correlation Coefficient (PCC), Variable Importance of Projection (VIP) and Function of Fitting Error (FFE) are used to measure the importance of each regression factor in every step. As a case study, the Taihu daily water level forecasting model has been build. The results show the following: The partial least squares regression can effectively overcome the serious multicollinearity among factors, and the regression coefficient accords with the hydrological law; The Q2cum score is closely associated with the accuracy of the forecast period, effectively reflect forecast ability of the model; The FFE plays a better role in measuring the importance of factors than VIP and PCC; Based on FFE, by forcing the impact of the factors to be continuous, the forecast ability of the model declines, but the model accords with the hydrological law much better, and is more practical.