Recurrent neural network based modelling and temperature control of a bioreactor for ethanol production

Abstract A bioreactor is a complex, nonlinear, and time-varying system that requires accurate mathematical models and superior control performance. Controlling the temperature has an enormous impact on the overall performance and efficiency of ethanol production. Data-driven modelling and deep learning techniques like recurrent neural networks capture temporal characteristics. Hence, in this study, a gated recurrent unit (GRU), a type of RNN is designed for the control of temperature in the bioreactor system based on a model-based control framework. The generated input-output data from the first principles model of the bioreactor is provided for the training of GRU network model. The GRU model is tuned by adjusting hyperparameters such as number of hidden layers, hidden units, and epochs for the minimization of the prediction error. Similarly, another GRU network is trained for use as a controller. To track the test setpoints and reject disturbances with minimum error, the controller network architecture is tuned. This procedure is repeated and compared with another popular RNN known as long short-term memory (LSTM) network as well as with bi-directional long short-term memory (Bi–LSTM) network also. Bi–LSTM based model shows the minimum root mean squared error (RMSE) i.e., 0.0076 whereas LSTM and GRU models shows 0.0085 and 0.0077 respectively. The servo and regulatory response of these network controllers are evaluated in terms of a standard performance measure such as integral squared error (ISE). The servo and regulatory response of these network controllers are evaluated in terms of a standard performance measure viz., integral squared error (ISE). In the servo control, the GRU based controller provides less ISE i.e., 37.02 whereas in LSTM, Bi–LSTM and IMC–PI based controller provides 39.66, 51.10 and 50.43 respectively. Similarly, in the case of regulatory response, the GRU based controller rejects the disturbance i.e., input flow rate effectively.