The detection of working fatigue has been widely studied but is mainly about detecting the state based on current indicators. If indicators can be predicted, future fatigue states can be forecasted based on existing methods, so measures can be taken in advance to better cope with working fatigue. Feedback neural networks have been successfully proven to perform well in sequential applications. To address the challenge of indicator prediction, a gated recurrent unit neural network-based method is proposed, which contains two phases: First, an overlapping window sampling data preprocessing method is adopted to increase diversity. Then, a gated recurrent unit neural network with hyperparameters optimized by Bayesian is used for prediction. Experimental results on the mean RR Interval prediction problem have shown that the proposed method can be successfully applied to the prediction problem.