Dynamic modelling of electrodialysis with bipolar membranes unit using NARX recurrent neural networks

Electrodialysis with bipolar membranes (EDBM) is an innovative and effective process for the simultaneous production of acid and base solutions from salty streams. It has been proven to play a key role in several circular economy approaches to valorize waste industrial brines, but it can also be used for in situ generation of chemicals, especially in remote areas. The adoption of such technology at industrial scale requires reliable modelling tools capable of predicting both dynamic and stationary operations as process conditions vary, such as energy supplied to the system and the target concentration of chemicals. In this study, nonlinear autoregressive models with exogenous inputs (NARX) were applied for the first time to EDBM to predict the behaviour of this complex and nonlinear process. Thus, an effective and low computational demanding neural-based modelling tool was developed. As a preliminary step, the network was trained with three different datasets, generated by a fully validated model. The best architecture was chosen to give good performance, testing the network with a new dataset. The NARX network accurately predicts the different behaviour of EDBM outputs (i.e. voltage and solutions conductivities) showing low average discrepancies between predicted and true values (lower than 0.5 %). These results suggest the possibility of using neural network-based models to effectively optimize and control EDBM process. Next step will focus on the training and validation of a network obtained with a set of data from a real EDBM plant.