基于SVR与特征变量选择方法的煤炭发热量预测

Prediction of coal calorific value based on SVR and characteristic variables selection method

目前应用较多的煤炭发热量预测模型主要以传统的线性回归模型为主,且大多以灰分、挥发分、固定碳等工业分析数据为主,而较少考虑元素分析指标对发热量预测效果的影响。因此,此类方法存在对煤的发热量预测精度低、适用范围窄的缺陷。本文基于我国不同地域的60组煤质分析样本,同时考虑工业分析和元素分析指标对发热量的影响,采用平均影响值法(Mean Impact Value,MIV)和粒子群算法(Particle Swarm Optimization,PSO)相结合的方法,对影响煤炭发热量的指标进行特征变量筛选,并结合支持向量回归方法(Support Vector Regression,SVR)对煤炭发热量进行非线性建模和预测。结果表明:对于我国不同地域的60组煤炭样本,煤的工业分析和元素分析指标中,仅灰分、碳含量与发热量之间存在一定的线性相关性,其余指标与发热量的线性相关性较差,煤的发热量预测应优先考虑非线性建模方法。经特征变量筛选后发现,煤的工业分析指标中水分、灰分和挥发分三者对发热量的影响均较大,而元素分析中仅碳含量对发热量影响较为显著;以影响显著的4个指标为特征变量进行发热量预测,SVR方法对煤发热量的预测精度显著提高,相关系数达0.9827%,远高于其他线性回归模型。同时,不同发热量模型预测结果也表明,煤炭样本来源地域越接近,各方法预测精度差异越小,应依据样本地域差异选取合适的预测方法。

Currently,most of the coal calorific value prediction models are based on the traditional linear regression model.Also,most of them are based on proximate analysis such as ash,volatile and fixed carbon,while less consideration is given to the effect of elemental analysis on calorific value prediction.Therefore,these methods have low accuracy and narrow application range for predicting coal calorific value.To overcome the shortcomings of traditional linear regression model in coal calorific value prediction,a nonlinear modeling and prediction method by support vector regression(SVR)coupled with mean impact value(MIV)was proposed.By using 60 sets of coal quality analysis sample data from China and considering the influence of proximate analysis and elemental analysis on calorific value,the characteristic variables of influencing the calorific value were selected.Based on this,the nonlinear predication model of coal calorific value was built.The results show that for 60 sets of coal quality analysis sample data from China,only ash and carbon content show a certain linear correlation with calorific value in the indexes of proximate analysis and elemental analysis,the other indexes show a poor linear correlation with calorific value.Thus,nonlinear modeling method should be given priority in coal calorific prediction.After the selection of the characteristic variables of influencing the calorific value,it is found that the influence of moisture,ash and volatile on calorific value is greater in the proximate analysis of coal,while the effect of only carbon content on calorific value is more significant in elemental analysis.When predicting calorific value with above four significant indexes as characteristic variables,the prediction accuracy of SVR method is improved significantly,and the correlation coefficient is 98.27%,which is much higher than other linear regression models.In addition,the prediction results of calorific value based on linear and nonlinear models show that the closer the coal sample source area is,the smaller the difference of prediction accuracy of each method is.It is appropriate to select different predict methods according to the regional differences of samples.