Adaptive on-line optimal control of bioreactors: application to anaerobic degradation

Abstract

The control of bioreactors is particularly important due to ever-changing reactor and feed conditions. In adaptive optimal control reactor, conditions are adjusted to obtain the best performance, as determined by a performance index (PI), which is often a function of more than one output variable. An empirical, linear input-output model is used whose parameters are updated at each sampling time using on-line measurements. In this example, for anaerobic wastewater treatment, two relations are used of the form, y 1( t) = b 1, 1 u( t−1 - d 1) + c 1; y 2( t) = b 2, 1 u( t − 1 − d 2) + c 2. These are linear models relating feed flow rate, u, to methane production rate, y 1, and organic acids concentration y 2. The performance index (PI) was taken as a compromise between high methane rate and low organic acids as PI = methane production rate - Constant X (Total organic acids conc.). This was expressed in terms of the model parameters and maximized by the method of steepest ascent. In this way the feed rate could be continually adjusted for changing conditions, as caused by feed disturbances or biological adaptation. This dynamic optimization method has considerable advantage for slow systems, in that a steady state is not needed. One and two stage, anaerobic biofilm fluidized sandbed reactors were used to continuously degrade whey to biogas and residual organic acids. A PC computer was used to handle the on-line GC data and to serve as controller for the feed flow rate. Preliminary work involved testing conventional feedback control, for which pH setpoint control performed best. A complex simulation model based on mechanistic, dynamic mass balance-kinetic relations was established to describe the experimental response of the reactors. The simulation model was used to design the controller and to determine useful values for the control parameters; it proved extremely useful as it made lengthy experiments unnecessary. The PI function was shown by simulations to have a suitably sharp optimum. Automatic control experiments for one and two stage reactors were made in which responses were obtained for step changes in the performance index constant and for square wave increases in the feed concentration. The optimizer performed well, changing the feed rate to maintain the loading approximately constant with changed feed concentration. For a single stage system a new optimum was found after 5 h for a change in the PI constant and after 2 h for a step change in the feed concentration.