Wafer processing cycle time is an important indicator in the production process of wafer fabrication systems, which can help to develop more reasonable processing scheduling and improve production efficiency. Problems such as the high dimensionality of wafer processing shop monitoring data and large redundancy among parameters make the process of building cycle time prediction models challenging. To address these problems, a hybrid feature extraction method based on relevant features with density peak clustering algorithm (DPC-Relief-F) is proposed in this paper, which can filter out the key feature subsets in the sample features, thus reducing the model input data dimension and improving the model training speed as well as the prediction accuracy. In addition, a parallel training prediction model combining fuzzy C-mean with back propagation network (FCM-BPN) is proposed to speed up the training process under large-scale data. Experimental results show that the model built by this method has higher prediction accuracy as well as shorter training time than traditional prediction models in the face of larger data.