On-line flame signal time series analysis for oil-fired burner optimization

Abstract

Flame monitoring and characterization have been recognized as an important aspect for burner combustion tuning. Flame flickering and its electromagnetic flame spectrum have been extensively demonstrated to provide characteristic information of flames in the combustion processes. This information can be used for combustion and emissions optimization and control. A system based on flame characterization would comprise of an on-line monitoring instrumentation, a time series processing algorithm and a control scheme. This paper reports the development of a smart software for flame data processing, to provide flame indexing and infer combustion stoichiometry under a range of combustion conditions. The software processes standard flame scanner data in the time and frequency domain, and employs standard statistical measures, shape factor, information entropy analysis and plant data for flame classification. Artificial probabilistic neural networks algorithms were used for clustering the flame spectral data. Probabilistic neural network is a supervised learning algorithm but includes no weights in its hidden layer. It is often an excellent pattern classifier that outperforms other classifiers, including back-propagation. Data from the scanner flame monitoring system, the plant data acquisition system and the stack emissions monitoring system are processed with proprietary algorithms to provide burner flame classification. Additionally, an expert system utilizes the flame classification information and interfaces with the plant operators for open-loop control of the burners and to achieve optimal scheduling of burner maintenance. The software was deployed at the Mexican Commission of Electricity’s CTPALM Unit 1. This unit consists of a multi-burner unit that fires Bunker C oil. The results demonstrated the benefits of the software.