Empirical mode decomposition with least square support vector machine model for river flow forecasting

Abstract

Accurate information on future river flow is a fundamental key for water resources planning, and management. Traditionally, single models have been introduced to predict the future value of river flow. However, single models may not be suitable to capture the nonlinear and non-stationary nature of the data. In this study, a three-step-prediction method based on Empirical Mode Decomposition (EMD), Kernel Principal Component Analysis (KPCA) and Least Square Support Vector Machine (LSSVM) model, referred to as EMD-KPCA-LSSVM is introduced. EMD is used to decompose the river flow data into several Intrinsic Mode Functions (IMFs) and residue. Then, KPCA is used to reduce the dimensionality of the dataset, which are then input into LSSVM for forecasting purposes. This study also presents comparison between the proposed model of EMD-KPCA-LSSVM with EMD-PCA-LSSVM, EMD-LSSVM, Benchmark EMD-LSSVM model proposed by previous researchers and few other benchmark models such as Single LSSVM and Support Vector Machine (SVM) model, EMD-SVM, PCA-LSSVM, and PCA-SVM. These models are ranked based on five statistical measures namely Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Correlation Coefficient ( r ), Correlation of Efficiency (CE) and Mean Absolute Percentage Error (MAPE). Then, the best ranked model is measured using Mean of Forecasting Error (MFE) to determine its under and over-predicted forecast rate. The results show that EMD-KPCA-LSSVM ranked first based on five measures for Muda, Selangor and Tualang Rivers. This model also indicates a small percentage of under-predicted values compared to the observed river flow values of 1.36%, 0.66%, 4.8% and 2.32% for Muda, Bernam, Selangor and Tualang Rivers, respectively. The study concludes by recommending the application of an EMD-based combined model particularly with kernel-based dimension reduction approach for river flow forecasting due to better prediction results and stability than those achieved from single models.