Anaerobic membrane bioreactors have become an environmentally friendly solution for wastewater treatment.
The lack of sufficiently accurate mathematical procedures to model their behaviour and the fouling process of
the membranes, poses a challenge when trying to optimise their energy consumption and maintenance costs.
An accurate model of the fouling process of the membranes is critical to make the most of this technology. This
is a perfect scenario in which to introduce neural networks (NN) as an alternative to mathematical modelling.
However, the duration of the experiments and the difficulties in measuring some relevant variables, make it
hard to collect high quality datasets to train the NN. Our goal is to obtain a good prediction of the fouling
status of the membranes to enable an adjustment of operation conditions and maintenance procedures ahead
in time. To do so we must obtain high quality datasets to train our neural networks. The combination of static
and dynamic networks enables us to leverage the best prediction capabilities of each one. This combination
requires a preprocessing of the datasets that separates trends from oscillations. The outputs obtained need to
be put together to build up the predicted evolution of fouling. Accurate predictions are then extended from
25 to up to 75 filtration cycles. To maintain and even extend accuracy after sudden changes in operating
conditions, retraining the NN every 25 cycles is proposed. AI based real time predictions open a new scope
for decision making, and optimisation in the field of anaerobic membrane reactors.
