Seasonal streamflow forecasts using mixture-kernel GPR and advanced methods of input variable selection

Gaussian Process Regression (GPR) is a new machine-learning method based on Bayesian theory and statistical learning theory. It provides a flexible framework for probabilistic regression and uncertainty estimation. The main effort in GPR modelling is determining the structure of the kernel function. As streamflow is composed of trend, period and random components. In this study, we constructed a mixture-kernel composed of squared exponential kernel, periodic kernel and a rational quadratic term to reflect different properties of streamflow time series to make streamflow forecasts. A relevant feature-selection wrapper algorithm was used, with a top-down search for relevant features by Random Forest, to offer a systematic factors analysis that can potentially affect basin streamflow predictability. Streamflow prediction is evaluated by putting emphasis on the degree of coincidence, the deviation on low flows, high flows and the error level. The objective of this study is to construct a seasonal streamflow forecasts model using mixture-kernel GPR and the advanced input variable selection method. Results show that the mixture-kernel GPR has good forecasting quality, and top importance predictors are streamflow at 12, 6, 5, 1, 11, 7, 8, 4 months ahead, Nino 1 + 2 at 11, 5, 12, 10 months ahead.