Applying and assessing multi-output support vector regression with rainfall as additional output for monthly river flow forecasting

Abstract

Developing data-driven models for river flow forecasting has received great attention in recent years. However, no study has applied multi-output support vector regression (MSVR) for river flow forecasting. In this paper, we presented a monthly river flow forecasting model using MSVR with both river flow and rainfall data as model output. By experiments on two station data of the National River Flow Archive, we compared the proposed MSVR model with five commonly used single-output forecasting models. Besides, potential influencing factors of the proposed model related to the training samples size, the kernel function type, input variable scenario, and scenario of output combinations were tested. Results in our experiments revealed three significant findings. Firstly, MSVR outperformed the five single-output models in all test data. Secondly, the additional output-rainfall in MSVR can be conducive to forecast river flow only under the configuration of appropriate input variables and kernel type. Thirdly, with the tiny training sample size, the proposed model showed significantly superior to support vector regression. Based on these evidences, it is suggested that through the linkage of the output modeled by MSVR, more beneficial information can be learned for forecasting river flow. Therefore, the proposed method had a promising potential for river flow forecasting.