(179b) A Sequence to Sequence Time Series Forecasting Model for Batch Processes with Seasonality

In chemical manufacturing, seasonality is a commonly faced problem, especially in pharmaceutical industries where a series of batch operations are conducted. Generally speaking, seasonality stems from equipment aging, periodic maintenance, catalyst decay, feed quality, etc. Process engineers are challenged to forecast such a seasonality lurking in the system so that they can become proactive in identifying and preventing potential problems. In current work, we demonstrate the adoption of the sequence to sequence learning framework [1] to overcome such a challenge. A time series forecasting model is built in Tensor Flow, and it is tested on the data generated from the benchmark Penicillin bio-reactor simulator. The model successfully predicts the cyclic change of penicillin production rate with the periodic decay of the biomass growth rate. Such results illustrate the successful application of Data Analytics to wrestle with real problems in chemical manufacturing within the Digital Transformation era.

[1] Sutskever, I., Vinyals, O., & Le, Q. V. (2014). Sequence to sequence learning with neural networks. In Advances in neural information processing systems (pp. 3104-3112).

[2] Birol, G., Ündey, C., Parulekar, S. J., & Çınar, A. (2002). A morphologically structured model for penicillin production. Biotechnology and bioengineering, 77(5), 538-552.