Research On The Optimization Of Coal-blending Coking Experiment And Neural Network Coke Quality Prediction Model

The properties of raw coal used in coal-blending were investigated and coal blending schemes were optimized in laboratory conditions. The effect of particle size distribution on coking performance and coke strength was studied by screening and crushing the raw coal. Basing on the principle of coal petrography coal blending, two key factors of the maximum vitrinite reflectance and reflectance distribution of coal vitrinite were selected to optimize coal blending. The coke’s quality can be improved by regulating the proportion of different metamorphic grade fat coals and adding bonding agent (pitch) and leaning agent (coke powder), which provides the theoretical basis for comprehensive utilization of coking coal and industrial production of high quality coke. The results show that selective crushing of coal has little effect on the coking rate, but can improve the cold properties (crushing resistance strength M13 and abrasion resistance strength M3) and thermal strength (particle coke reactivity toward CO2 CRI* and strength after reaction CSR*) of coke; the maximum vitrinite reflectance and the reflectance distribution of blended coke are two effective controlled parameters for optimizing the coal blending. The CRI* of coke are decreased 1.95% and the CSR* are increased 3.45%, respectively by regulating original coal blending scheme, the thermal strength of coke is significantly enhanced; Cold properties of coke can be remarkably raised by addition of fat coal with lower expansibility and more plastic mass, the suitability among various kinds of coal is also existed, CSR* is increased 4.04% when LF(2) fat coal is replaced by 7% MF fat coal used, the effect of pitch on the quality of high binding property blending-coal coke is limited, but cold properties of coke is significantly improved by addition of 1% pitch and coke powder respectively.BP neural network is used to establish the coke quality prediction model, which contains three layers,5 input and 4 output variable. K-fold cross-validation algorithm is introduced to optimize the network, improve generalization ability and reduce the errors. The results show that the relative errors of M13, M3, CRI*, CSR* are in the range of 0.61～1.95%,2.31～9.73%, 0.05～6.07%,0.13～3.22%, respectively, which meet the requirement of design error.