Typical condition library construction for the development of data-driven models in power plants

Abstract

The performance of data-driven models to describe processes in thermal power plants highly depends on the operating data that are used for model training. Models developed using representative operating data with rich process information are more likely to produce accurate predictions. This paper presents a concept of typical condition library to collect informative operating data from power plants. First, a quantitative criterion to measure the condition information of operating data is defined by three factors including variation span, distribution status, and redundancy. Additionally, an analytic expression is provided to describe the condition information metric. Then, the genetic algorithm is used to search operating data samples with the largest information for the condition library construction. A numerical case is given to validate the proposed metric and the sample selection strategy. Finally, the typical condition library is applied to develop a model for a selective catalytic reduction (SCR) system in a coal-fired power plant. The results indicate that the SCR model trained using samples in the library can give accurate predictions, and the typical condition library is effective for initial data preparation of model development.