An apparent disadvantage of an Extreme Learning Machine compared to an Artificial Neural Network, though its training times for similar architectures is less by two orders of magnitude, is that the absence of an iterative minimization of error precludes any obvious mechanism for preventing overfitting. However, this mechanism can be introduced by $($choose the most realistic option$)$:

a$)$modulating the number of training samples for a given ELM architecture so that the bias$-$variance tradeoff settles at just right

b$)$for a given training data set which you have further split into training and validation data, try out a series of ELMs with different number of hidden layer nodes in each case, and then choose that which gives minimum validation error

c$)$modulate the random weights in the first hidden layer till you get the minimum validation error

d$)$experiment with the activation function in the hidden layer nodes till you converge on that which gives minimum validation error.