Combining forecasts for natural streamflow prediction

This paper proposes an approach to combine forecasts generated by a set of individual forecasting models in a simple and effective way. In principle, combination can be done using appropriate aggregation operators, but here we use a neural network trained with the gradient algorithm. The aim is to combine the forecasts generated by the different forecasting models as an attempt to capture the contributions of the most important prediction features of each individual model at each prediction step. The approach is used for streamflow time series prediction choosing, as individual forecasting models, periodic autoregressive moving average model (PARMA), and two fuzzy clustering-based forecasting models. Experimental results with actual streamflow data show that the combination approach performs better than each of the individual forecasting models and yet, when compared to a fuzzy neural network (FNN) evaluation, the suggested combination model shows lower prediction errors.