Forecasting Membrane Fouling in Filtration Processes using Univariate Data-Driven Models

Filtration processes play a crucial role in modern food production. Filtration techniques are, for example, used to purify wine or to produce protein powder out of high viscose suspensions. Characteristically, the filtration membrane becomes increasingly clogged during filtration by deposits of non-permeable particles on the surface. The filtration quality deteriorates caused by this membrane fouling until the process comes to a standstill due to a wholly blocked membrane. A forecast of the membrane fouling process allows appropriate actions to be taken early during filtration to maintain the process quality and reduce membrane blocking. This paper investigates if and how membrane fouling can be forecasted over time in filtration processes using non-phenomenological data-driven univariate stochastic or machine learning models. Prophet, ARIMA and LSTM models for time series forecasts are selected and evaluated on industrial data sets collected from Dynamic Crossflow Filtration plants with different observation lengths and forecast horizons. ARIMA and LSTM models show suitable results on the used data set and are evaluated with varying combinations of past observations and forecast horizons.