An L1-Regression Random Forests Method for Forecasting of Hoa Binh Reservoir´s Incoming Flow

Random Forests (RF) method has been widely used as a powerful ensemble learning tool for forecasting problems. RF uses the least squares criteria to search the best split when growing trees and takes the mean over all trees to aggregate the final forecast. The performance may not be accurate when applied to data set with respect to the presence of outliers and skewed distributions. In this paper, we proposed to use the l1-norm as the splitting rule for growing trees and take the median to obtain the forecast values in the forest. The proposed RF is applied to forecast the incoming flow of Hoa Binh´s reservoir for 10 lead days. Experimental result showed that the proposed RF outperforms other state-of-the-art methods in reducing of RMSE measure, the proposed approach provides an useful and feasible method for forecasting the incoming flow problem.