Abstract: Linear models such as least squares (LS) extrapolation and autoregressive (AR) models are used to predict Earth's polar motion. However, polar motion predictions obtained by linear models are not fully satisfactory owing to the non-linear and non-stationary variations in polar motion. In this paper, the extreme learning machine (ELM), a new training algorithm for single hidden layer feedback neural network (SLFN), is employed to predict medium-and long-term polar motion. The LS extrapolation model is first employed to separate a linear trend, Chandler and annual wobbles (CW and AW) from observed polar motion time-series. The extracted linear trend, CW and AW are then predicted by using the above-mentioned LS extrapolation model. Secondly, the ELM is used for modeling and forecasting the residual time-series obtained by subtracting the LS extrapolation curve from the original polar motion data. In ELM training some special issues such as generation of training patterns and topology of the network for enhanced medium-and long-term prediction of polar motion are taken into consideration. The subsequently forecasted residuals are then added to the LS extrapolation so as to obtain the predicted values of polar motion. In order to demonstrate the effectiveness of the presented method, the results of the predictions of polar motion are analyzed and compared with those obtained by the back propagation (BP) neural network and Earth Orientation Parameters Prediction Comparison Campaign (EOP PCC). The results show that the predictions are comparable with or better than those obtained by other existing prediction methods and techniques in terms of the mean absolute error (MAE).