Nonlinear System Identification Based on Ensemble Learning: The Cascaded Tanks System Case

Ensemble approach plays an important role in the currently scientific scenario, once it can be used to improve the accuracy of time series models. This approach combines several weak models to build an efficient forecasting model able to deal with nonlinear systems. In this aspect, the aim of this paper is to propose an ensemble approach based on stacking to make predictions k-step-ahead for a cascaded tank system benchmark problem. Base learners (BL) such as partial least squares, support vector machines for regression, boosted generalized linear and weighted k-nearest neighbors models are adopted as well as boosted linear model is used as meta-learner. The performance of proposed ensemble is compared with BL performance by root mean squared error, determination coefficient and mean absolute error criteria. Results show that performance of proposed approach are better than the BL performance. Indeed, satisfactory forecasting results regarding to accuracy are found.