Prediction algorithm is often adopted to eliminate the impact of controller delay on the compensation effect in the digitalized control system. For slowly changed loads， the current is usually easy to predict， but for dynamic impact loads， the current prediction is very difficult. Very few existing prediction algorithms are satisfactory， and may even lead to serious deterioration of the compensation effect. In this paper， based on the adaptive filter， a dynamic prediction algorithm of harmonic current is proposed for the dynamic impact load such as rolling mill. A non-linear load current signal is fitted with a polynomial which is superimposed of a series of harmonic current signals， and the coefficients of the polynomial is real-time estimated with the recursive least square algorithm with forgetting factors. Simulation results show that the presented prediction algorithm is easily implemented in the digitalized controller because of its moderate computational complexity， and can accurately predict the dynamic impact load current to greatly improve the controller's compensation effect.