Application of Neural Network Combined With CFD Modeling and Combustion Tuning in Large Coal-Fired Boilers

Abstract

The objective of the present work was to develop an optimization method for the prediction of the behavior of coals or coal blends in utility boilers, in order to specify the performance and pollutant emissions during the firing. Two methods have been used to study the performance of single coals or coal blends in power station boilers (1) experimental tests, where the coal/blend was fired in either a power station or in a test rig, and (2) use of coal combustion computational fluid dynamic (CFD). Here we will discuss both methods. We present experimental results, for 575 MWe tangentially-fired Combustion Engineering boilers of Israel Electric Corporation and 50 kWth test rig of Ben-Gurion University, that show the control of NOx and carbon content in fly ash (LOI). In addition to the experimental measurements we also established a large data base using a CFD code for a large spectrum of operational conditions. Validation of CFD results was made by comparison with both test rig and full-scale boilers measurements. Only after ensuring that good fit was obtained between experimental measurements and CFD results, was CFD used to establish the data base for coals/blends at a large spectrum of operational conditions. In some cases CFD was run for coals/blends never burned in the boiler, but burned in the test rig. The data obtained, experimental, showed that with tuning and modified nozzles NOx was considerably reduced: from 1200 to 570 mg/dNm3 @ 6% O2 for South African coal at full load. At partial loads NOx emission dropped from 1400 to about 800 mg/dNm3 @ 6% O2. High volatile coals, such as Colombian and Indonesian, firing led to additional NOx reduction to around 400 mg/dNm3 @ 6% O2 at full load. A very large data base was obtained in this effort and brought us to the idea of extending it by using a neural network algorithm [1]. We used these data as a base for the development of a code based on neural network and a mathematical optimization algorithm. The code was primarily intended for use by the plant personnel for better tuning coal-fired boilers to reduce NOx and minimize heat rate. The neural network develops non-linear mapping functions between the outputs of NOx, heat rate, LOI, etc. and the controllable boiler input parameters. The mapping functions are then analyzed by the mathematical optimization algorithm and optimal boiler operating condition are identified. Further, based on networks and a mathematical optimization algorithm we found a proper Adaro and KPC (Indonesian coals) blend and operation condition that led to NOx emission reduction less than 400 mg/dnm3 in a 575 MWe tangentially firing unit with a conventional firing system. This result was verified in experimentally in the boiler. The results presented in this work clearly show that the developed method for reduction emission and performance optimization is available and capable to achieve operational or environmental goals.