Research On Coal Pyrolysis Characteristics And Coal Ash Melting Characteristics Based On BP Neural Network

The efficient clean utilization of coal has become an inevitable trend,based on the requirements of environmental protection and the characteristics of China's energy structure.However,the thermal transformation methods such as coal pyrolysis and coal gasification are often carried out under the harsh conditions?high pressure and temperature?.Meanwhile,due to the complex structure of coal and the contributing factors,it is difficult to summarize the reaction law.The BP typical network in artificial neural network has strong nonlinear mapping ability as well as flaws of slow convergence speed and falling into local minima.Therefore,a variety of optimization algorithms combined with stepwise predictive verification are used to find out the improved BP neural network with excellent performance and we apply it to develop coal pyrolysis prediction model and coal ash melting temperature prediction model to provide reference to the process of optimization and design in coal thermal conversion process.In the coal pyrolysis prediction model,five factors including pyrolysis temperature?T?,volatiles(V_daf),ash?A_d?,the ratio of carbon to hydrogen?C/H?and the ratio of oxygen to hydrogen?O/H?are selected as input variables,the percentage of residual mass of coal was chosen as the output node of the network,and the 5-11-1 model is established.Firstly,the predict ability of different excitation function combinations is explored by this model.The results show that when the linear function?Pur?is selected as the excitation function in the hidden layer,the network does not converge regardless what excitation function is selected in the output layer.When the excitation function is nonlinear function?Bip-S,Log-S,Sin,Tan-S?in hidden layer and Pur in output layer selected,the networks show a good prediction accuracy which also means mean square error of each function combination is of same magnitude.It can be seen that the BP performs better with hidden layer choosing the nonlinear function and the output function selecting Pur.Based on the excitation function combination,performance of different optimization techniques including additional momentum and adaptive learning rate?A-BP?,genetic algorithm?GA-BP?,particle swarm optimization algorithm?PSO-BP?and hybrid algorithms?PSO-GA-A-BP?,was further explored.The four improved neural networks are advantageous in terms of training error and convergence speed compared with ordinary BP model,namely,the improved algorithm plays an active role in the operation of the network.However,different optimization methods impose different effects on the BP neural network.The A-BP model can make the network converge quickly with poor stability,which is mainly due to the change of the learning rate itself;GA-BP and PSO-BP models have improved the convergence rate at the beginning of the network operation,which also confirms GA-BP,PSO-BP is the design for initial weight and threshold;PSO-GA-A-BP model not only has the fast convergence ability,but also ensure the stability of the network in the running process.It can be seen that the PSO-GA-A-BP model has most excellent performance compared with other improved neural networks.In addition,we conclude that the convergence rate of each model is PSO-GA-A-BP>A-BP>GA-BP>PSO-BP>BP.The PSO-GA-A-BP neural network was selected to explore the prediction effect.The results show that the predicted datas is highly consistent with the experimental datas and superior precision as the relative error between the predicted value and the experimental value is under 3.5%,which shows the model has a good practicality in the field of coal pyrolysis.In order to verify the stability of the network output results,we conduct nine parallel experiments and the mean square error of nine parallel experiments in the range of0.0023-0.0057,the number of operations in the range of 4715-6931.It can be seen that the mean square error of the output results of the model is not only in the same order of magnitude and very close,which directly shows that the stability of the output results of the model;the number of runs of the model is also on an order of magnitude and very close also shows that the network has good stability.In the coal ash fusing characteristic prediction model,the model of the 1-4-4 structure is established that the network chooses different coal blending ratio as the input factor,and the four characteristic temperature of the ash melting temperature are used as output factors,to explore the influence of the blending proportion between Jincheng coal?JC?and Xiangyang coal?XY?on the fusing characteristics of coal ash.The results show that the model not only predicts the trend that the four characteristic temperatures of coal ash fusing temperature change with the proportion of coal blending,but also finds that the ratio of entering the platform area is 24%and the proportion out of the platform is 42%.This prediction conclusion is fully verified in the further coal refining experiment.In addition,the prediction accuracy of the prediction model is more accurate than that of the previous empirical formula.In the prediction model of influence of CaO,Fe₂O₃ proportion on FT,the model of structure 2-4-1 is established that CaO and Fe₂O₃ percentage are chosen as the input factors,the FT as the output,to explore the influence of CaO and Fe₂O₃ on FT.The results show that CaO and Fe₂O₃ can reduce the FT of coal ash.With increase in CaO content from 1.2%to2.4%,FT shows a continuous significant decrease regardless whether Fe₂O₃ content increases or not.FT decreases significantly with the increase of Fe₂O₃ content within 1.8%;change of FT is gentle with the increase of Fe₂O₃ content beyond 1.8%.It can be deduced that the order of action of reducing coal ash melting temperature is CaO>Fe₂O₃,in addition,the results of XRD analysis show that the calcium-containing aluminosilicate is significantly higher than that of Fe-containing aluminosilicate which further supporting this conclusion.