Abstract:
Compared with continuous process, batch process is an important industrial production mode with more flexibility. Products from each batch depend on market demands, which will make the process switch between different product stations. So, the final quality of products may become unstable due to various composition of raw materials. This kind of uncertainty brings great challenge to process modeling. Aiming at the problem of raw material uncertainty and the requirement of actual predicted value, the loss function of bidirectional gated recurrent unit neural network is modified, and then, utilized to predict the final quality prediction of unequal-length batch processes. The bidirectional gated recurrent unit neural network can integrate the time series information of the batch data, and fully exploit the inter-batch timing characteristics caused by the uncertainty of the raw materials. The improved loss function imposes different penalties on different predicted values such that the predicted values can meet the requirements in actual production. Finally, the proposed method is compared with multiple partial least squares (MPLS), neural network (NN), support vector regression (SVR) and gated recurrent unit (GRU) via an industrial process. It is shown via the results that the proposed method has better applicability and performs higher accuracy.