Run-to-run Trajectory Prediction of Uneven-length Batch Processes Using DTW-LSTM

This paper handles with the problem of the run-to-run trajectory prediction of batch processes with uneven batch length. Most current data-driven works focus on the run-to-run variations during both batch trajectory modeling and prediction stages. However, batch-to-batch correlations should be drawn extreme attentions to when gradual changes exist in batch sequence. To obtain a better batch trajectory prediction performance of uneven-length batch processes, dynamic time warping (DTW) and long-short term memory (LSTM) neural network are introduced in this work to extract batch-to-batch correlations. Firstly, the recursive DTW is used to synchronize uneven batch samples. Then, the LSTM neural network is introduced to extract the run-to-run batch correlations during the trajectory modeling stage. Finally, online batch trajectory prediction can be implemented according to the offline LSTM model. A simulation based on the fed-batch penicillin fermentation process is provided to testify the effectiveness of the proposed method.