A Dynamic ELM with Balanced Variance and Bias for Long-Term Online Prediction

Abstract

For long-term online prediction of nonlinear time series, how to determine a feasible network architecture that conforms to the time-varying data stream is recognized to be a challenging problem. To deal with the issue, a dynamic ELM with balanced variance and bias has been proposed. A suitable fitting degree, which contains model applicability at sequential learning phase, is taken into consideration. Based on the shifting error of the sequence fragment, the automatic model update strategy is exploited. Transformable parameters help reduce the overfitting and underfitting at the same time, and avoid the trial and error caused by user intervention effectively, so as to guarantee the feasibility for long-term online prediction. Furthermore, hidden node number and the regularization parameter can be calculated according to the fast-changing test data, thus building an optimum network architecture quantificationally. Experimental results verify that the proposed algorithm has better generalization performance on various long-term regression problems.