A hybrid support vector regression framework for streamflow forecast

Abstract

Abstract Monthly streamflow time series are highly non-linear. How to improve forecast accuracy is a great challenge in hydrological studies. A lot of research has been conducted to address the streamflow forecasting problem, however, few methods are developed to make a systematic research. The objective of this study is to understand the underlying trend of streamflow so that a regression model can be developed to forecast the flow volume. In this paper, a hybrid streamflow forecast framework is proposed that integrates factor analysis, time series decomposition, data regression, and error suppression. Correlation coefficients between the current streamflow and the streamflow with lags are analyzed using autocorrelation function (ACF), partial autocorrelation function (PACF), and grey correlation analysis (GCA). Support vector regression (SVR) and generalized regression neural network (GRNN) models are integrated with seasonal and trend decomposition to make monthly streamflow forecast. Auto-regression and multi-model combination error correction methods are used to ensure the accuracy. In our experiments, the proposed method is compared with a stochastic autoregressive integrated moving average (ARIMA) streamflow forecast model. Fourteen models are developed, and the monthly streamflow data of Shigu and Xiangjiaba, China from 1961 to 2009 are used to evaluate our proposed method. Our results demonstrate that the integrated model of grey correlation analysis, Seasonal-Trend Decomposition Procedure Based on Loess (STL), Support Vector Regression (GCA-STL-SVR) exhibits an improved performance for monthly streamflow forecast. The average error of the proposed model is reduced to less than one-tenth in contrast to the state-of-the-art method and the standard deviation is also reduced by more than 30%, which implies a greater consistency.