A probabilistic prediction model based on stacking ensemble learning for completion time in flexible shop-floor

Abstract

Completion time prediction is crucial for analyzing and monitoring the execution of shop-floor’s production planning. However, predicting completion time is still challenging because production process has high uncertainty and influenced by the interaction of various factors. To address above challenges, the stacking ensemble learning based probabilistic prediction model (SEL-PP) to predict completion time is developed. Thereinto, fully connected neural network (FCNN), random forest (RF) and gradient boosted regression tree (GBRT) are used as the first-level base learner to exhibit better nonlinear characteristics, and quantile regression neural network (QRNN) is used as the second-level meta-learner to rectify the mistakes in the prediction of the base learner. After that, the kernel density estimation (KDE) is employed for achieving probability density prediction of completion time. Moreover, particle swarm optimization (PSO) is adopted for key parameter optimization of SEL-PP. Taking the aircraft overhaul shop-floor as an example, a case study is constructed to demonstrate the feasibility and effectiveness of SEL-PP. Through comparing the results, it indicates that the SEL-PP model has better performance in predicting completion time of flexible shop-floor.