A Chemometric Approach to Process Monitoring and Control - With Applications to Wastewater Treatment Operation

Abstract: In this work, various aspects of multivariate monitoring and control of wastewater treatment operation are discussed. A number of important difficulties face operators and process engineers when handling online measurements from wastewater treatment processes. These include, for instance, a high number of correlated measurement variables, non-stationarities, nonlinearities and multiscale process behaviour. A systematic way to handle and analyse data is needed to effectively extract relevant information for monitoring and control. In this work, a chemometric approach is taken. Principal component analysis (PCA) is used to reduce both the dimensionality of the problem and the noise level in data. However, it is shown that basic PCA is not sufficient to describe the process adequately. There are mainly two reasons for this. First, the processes display a non-stationary behaviour due to the diurnal, weekly and seasonal variations in the composition of the wastewater. Second, disturbances and events occur at different time scales making basic PCA less suitable. The problem of non-stationary data is overcome using adaptive PCA in terms of updating of the scale parameters as well as the covariance structure. It is shown that adaptive PCA significantly improves the monitoring results as the model adapts to new process conditions without losing its ability to detect deviating process behaviour. To solve the problem of disturbances that occur at different time scales multiscale PCA is used. Multiscale PCA is a combination of multiresolution analysis and PCA. Measurement signals are decomposed into several time scales, and PCA models at each scale are identified. By doing so, the sensitivity to small process deviations that otherwise are obstructed by the diurnal variation is considerably increased. By omitting the lowest time scale from the analysis, the remaining time scales will inherently be (practically) stationary since this corresponds to using a highpass filtered version of the data. Another solution, where the PCA models at each scale are made adaptive is also presented. Using the monitoring results to adjust the process in a supervisory control manner is discussed. Two different methods are presented. The first is based on a multistep procedure. The current operational state is detected and classified using clustering in the principal component space. This information is used to determine appropriate setpoints for local controllers so that the process returns to what is considered normal operation. In the setpoint determination step, both static and dynamic models are used. The dynamic models are used within the framework of model predictive control (MPC). The multistep approach is best suited for extreme event control, since nonlinear and discrete control actions easily can be incorporated. The second method to integrate monitoring and control is based on PCA. Here, the inverse PCA model is used to directly calculate appropriate setpoints for the local controllers so that the process can be controlled to attain specified output requirements. The controller can be seen as a multivariate feedback controller implemented on top of the local control system. It is shown by simulation studies that both methods for supervisory control can successfully be used to control the process according to the control objectives.