Abstract: In order to overcome the disadvantage of traditional feedforward neural networks in long-term prediction of machinery condition, a new neural model, so called the multi-step recurrent prediction model based on recurrent neural network is proposed. A learning algorithm of recurrent model for long-term prediction is also presented, which is supposed to obtain better predictions of machinery condition. The feasibility of recurrent neural model is examined by applying it to forecast a simulated time series and predict the behavior of large rotating machinery. Experimental results revealed that the recurrent model could achieve better prediction accuracy and provided cogent proof for realization of prognostic maintenance.