SUMMARY
The eutrophication problems in rivers, lakes and coastal waters, have increased the 
requirements for nutrient removal from wastewaters before to their discharge into 
natural water bodies. Initially, the organic matter and nitrogen was usually removed 
by biological processes while the phosphorus by means of chemical processes. However, 
the advantages of biological phosphorus removal have led this process to be gradually 
implemented in many wastewater treatment plants (WWTPs).
The control of a complex process as the biological wastewater treatment including 
biological phosphorus removal requires the measurement of those quality variables 
which are key indicators of process efficiency (ortophosphate concentration, ammonium 
concentration, chemical oxygen demand,...). Methods to on-line measure these variables 
involve high investments and important maintenance costs. On the other hand, there are 
other variables (process variables: pH, conductivity, ...) which can be measured on-line 
by means of inexpensive, robust and low maintenance sensors, but they do not directly 
supply information on process performance. The large number of on-line process variables 
collected nowadays at modern WWTPs, require structured techniques to extract the information 
from the huge amount of available data.
The main aim of this thesis has been to study the possibilities of using process variables 
to get information on quality variables as well as on process behaviour over time, by means 
of multivariate statistical methods. In this way, it is intended to detect any kind of anomaly 
which could occur (special cause of variation which can led to poor process performance or 
unusual process behaviour and/or poor effluent quality) and help in the isolation of its origin, 
thus, avoiding the use of quality probes whose cost make them unaffordable for most WWTPs.
This work has been focused in the study of a laboratory scale sequencing batch reactor (SBR) 
operated for biological phosphorus removal from wastewater. In this process, alternating 
anaerobic and aerobic conditions are imposed in order to favour the growth of polyphosphate 
accumulating organisms.
 

To analyse completed batches (off-line analysis), two approaches have been studied and 
compared: Nomikos and MacGregor (1995), and Wold el al. (1998). Due to the
limitations found in the observation level of the second approach for the data set Analysed, 
two modifications of the original methodology have been proposed to improve it: the first 
consists in replacing the partialleast squares (PLS) by a principal component analysis (PCA), 
and the second consists in using a different data preprocessing and replacing the PLS by a 
PCA to extract relevant information from the process.
For on-line batch process monitoring six strategies have been studied and compared: the 
observation level in the Wold el al. (1998) approach, the proposed alternative in which a 
different data preprocessing is used, the building of multiple reference models (evolving model), 
and the three imputation methods suggested by Nomikos and MacGregor (1995): zero deviation, 
constant deviation and missing data.
in order to predict the quality variables (phosphorus, potassium and magnesium), determined 
by means of laboratory analyses, several predictive models (soft sensors) have been developed, 
including models based on multivariate projection
techniques as well as artificial neural networks. AII the models have been validated ¡ing a 
data set which was not used for model building, and their predictive capability has been 
compared in terms of the mean squared error of the validation data set.
Fom the results obtained, an integrated system for on-line monitoring, diagnosis and prediction 
of the SBR process has been proposed. In this system multivariate process control methods based 
on projection to latent structures are used. Due to the fact that in the batch process analysed 
not only there are autocorrelations and cross¬-correlations among variables over time within a 
batch, but also important correlation between batches, the monitoring scheme proposed includes 
two levels. This allows monitoring new batches as they are evolving and also the overall process 
evolution associated with finished batches.
 

The results obtained in this thesis show that the application of advanced statistical techniques 
can help to achieve a more efficient and safe process operation using the information contained 
in the easy-to-measure process variables.