In this paper, the prediction effects of different BP neural network model were analyzed and the main factors affecting the prediction effects were also studied. These factors included network structure, learning sample quantities, hidden nodes, learning accuracy and etc. Based on the above analyses and studies, BP neural networks were built to predict characteristics, such as low heat and others, of blended coals, and the prediction accuracy is extremely high. The errors of the prediction cases were all less than 0.12%, and most of them were around 0.04%. In addition, the optimization of coal blending for a 125MW unit at a coal-fired power station was conducted with exhaustive method, and it is very directive to the practical coal blending. The characteristics of neural network are with close relation with the data extension of the input and output samples. The smaller the input and output data extension, the better the convergence of the neural network and the error of the network will be smaller. The structure of the neural network constructed in this paper is decided by structure of the input and output data, so it is universal and has strong expansion.
