The prevention of the phenomena of corrossion in the components of the water-steam cycle in a fossil power plant can save very important costs caused by unavailabilities and corrective maintenance. Because of this importance, power plants are equipped with chemical laboratories in charge of the monitoring of water purity. This monitoring is mainly based on the information coming from sensors able to measure the most representative properties of the water in the water-steam cycle. However, the important load of work to analyze these data every day added to the daily tasks to do, prevents the possibility of a thorough analysis of all the information available. An expert system SEQA (these initials in Spanish stand for Water Chemistry Expert System) was developed and installed in a power plant for the monitorization and diagnosis of the main properties of the water. The expert system was conceived as an important tool to help the personnel in the chemical laboratory and especially those in charge of the operation who are less acquainted with chemical terms. This expert system consists of several modules, one of which is the detector of the anomalous condition in the water. This module includes models based on neural networks characterizing the relationships existing among the dynamic evolution of the main variables involved. In this paper we will present the way to obtain models to characterize the relationships of the chemical variables observed in any operation condition in the absence of abnormalities. These models, a reference for the prediction of normal behaviour, can be used for the early detection of abnormalities in the evolution of the chemical variables. When a problem exists, the difference between relationships expected and observed can be used to explain the root cause of the problem. This is a very important advantage in respect to the usual existing instruments for chemical monitoring. The models are designed to run on-line in the plant. Some of the applied connectionist models, their ability to discover anomalies and a discussion of their implantation in a power plant will be included in this paper.
Keywords: Neural Networs, Diagnosis, Power Plant monitoring
PowerGen Europe'96, Budapest, (Hungry). June 1996
Published: June 1996.