An K-ELM Based Kiln Coal Feeding Trend Prediction Method

Rotary kiln is an important thermal equipment for the industrial production ofaluminum oxide,which has a large number of operating parameters and thecharacteristics of large time delay,nonlinearity,strong coupling and multipleinterference.The sintering of rotary kiln is a key procedure in the whole rotary kilnsystem.Because the coal feeding of the kiln is greatly influenced by the sintering kilnconditions,slurry composition and coal quality during the sintering process,even forthe experienced workers are difficult to judge the trend of increasing or decresing thecoal feeding in such a complex environment.In the situation of difficult to grasp thetrend of coal feeding in an artificial menthod,if a model used can accurately predictthe trend of coal feeding,it can be possible to stabilize the furnace condition andincrease the output and quality of clinker.According to this idea,the thesis take the amount of feed coal as the main objectof study,extract the trend features of the thermal data which base on the timeseries,research the relationship between the features of the thermal data and the trendof coal feeding,and propose a model based on kernel extremle learning ma-chine(K-ELM) to predict the trend of the amount of feed coal.For dealing with the kiln thermal data,the key point method based on increasingor decreasing the coal feeding was used to piecewise linear compression in the wholetime series,and then the change trend of each a piece data was extracted to eatablishthe trend of coal feeding data which use for training and testing.When the generalneural network was used for traning,it was affected seriously by the characteristics ofthe furnace thermal data which take with large noise and a lot of interference fac-tors,and it would have a low prediction accuracy.So, the K-ELM neural networkmodel with robustness was introduced in this article. Several neural network modelswere compared under the condition of several groups of standard data sets withnoise,and the results was also analysed.A fixed algorithm has not yet formed to select the parameters of the K-ELM inthe present research data.The cross validation method was adopted in this paper forthe parameters optimization which avoids the over-studing of neural network.The da-tabase of the coal feeding trend extracted in the step before was used to train and testmodels of ELM, SVM and K-ELM,and the results was analysed.The results of two experiments show that the model based on K-ELM has stronger anti-noise ability and the short-training time characteristic which can make aaccurate judgment for the coal feeding trends in a complex and noisy environment.This model was used to forecast the coal feeding trend more than5hours in an alu-mina oxide rotary kiln,and it’s about80%accuracy.The result proves that the methodused in the thesis is availabile.