Research on real-time prediction of completion time based on AE-CNN-LSTM

Abstract

For the flexible job shop (FJSP) completion time prediction problem considering environmental noise, the mathematical model of the problem, the sample generator and the completion time prediction model based on AE-CNN-LSTM are constructed. Firstly, to address the problem that job shop data is not easy to obtain, the sample generator is designed, including three parts: random use case generation based on hyperparameters, data annotation based on simulation scheduling algorithm, and dual-strategy data augmentation; then, multi-channel CNN and AE are used to extract features for the matrix state information and general state information, respectively; finally, the real-time and historical state features are passed into the LSTM to realize the prediction of completion time for the FJSP scheduling process under the fixed strategy. In the validation section, simulations and examples are conducted to verify the data enhancement effect, model training process, AE feature reconstruction, LSTM prediction accuracy, etc. The results show that the proposed sample generator can effectively generate and expand FJSP scheduling instances, and under different job shop scenarios, the AE-CNN-LSTM prediction model can efficiently extract the job shop state features, so as to maintain the average prediction accuracy of completion time at more than 90%.