Combustion optimization of a coal-fired boiler with double linear fast learning network

Fast learning network (FLN) is a novel double parallel forward neural network, which proves to be a very good machine learning tool. However, some randomly initialed weights and biases may be non-optimal performance parameters. Therefore, for the problem, this paper proposes a double linear fast learning network (DLFLN), in which all weights and biases are divided into two parts and each part is determined by least squared method. DLFLN is employed to model the combustion characteristics of a 330 MW coal-fired boiler and is combined with an optimization algorithm to tune the operating parameters of the boiler to achieve the combustion optimization objective. Experimental results show that, compared with extreme learning machine and FLN, although the DLFLN is assigned much less hidden neural nodes, the DLFLN could achieve much better generalization performance and stability under various operational conditions; in addition, the effect of the combustion optimization is very satisfactory.