ELM Variants Comparison on Applications of Time Series Data Forecasting

Extreme learning machine (ELM) which belongs to randomized algorithm categories, is versatile and an emerging learning algorithm. ELM has been developed for different application starting from pattern recognition, function estimation, regression analysis, time series analysis, and big data analysis etc. Unlike feed forward neural networks where slow convergence rate, imprecise learning parameters, presence of local minima are major bottles neck, This paper addresses these problems using different variants of ELM on some bench mark time series data. ELM and its variants where hidden nodes parameters like weights and biases are randomly generated and fixed during the time of learning process, also give results of weights as an output of single hidden layer feed forward neural networks (SLFNs) analytically. The paper performs experiments on two time series data and demonstrates that variants of ELM delivers good performance in generalized manner in several cases without compromising on accuracy.