Effects of Training Patterns on Predictions of Variations of Length Of Day Using an Extreme Learning Machine Neural Network

In this paper we investigate effects of training patterns on predictions of varations of Lenth Of Day (LOD) by an Extreme Learning Machine (ELM) neural network. We first discuss three types of training patterns, which are named as interval patterns, continuous patterns, and iterative patterns, respectively. We then present comparisons of the accuracies of the predictions using training patterns of the three types.We have found that results of using interval patterns are the worst; using iterative patterns can yield more accurate short-term predictions (of 1 day to 30 days away) as compared to using continuous patterns, but the situation is reversed in medium-term predictions (of 30 days to 360 days away). As for compution efficiencies, using iterative patterns is noticeably faster than using either of the other two patterns. Iterative patterns are thus suitable for real-time predictions of varations of LOD, while continuous patterns are suitable for medium-term predictions of varations of LOD. Efficiencies and accuracies are the primary concerns in these two types of predictions, respectively.