COMPARISON OF NEURAL NETWORK AND SUPPORT VECTOR REGRESSION MODELING OF HEAD TEMPERATURE OF DRY ROTARY CEMENT PLANT KILN, Abdolhossein Khosrozade, Nasir Mehranbod

Head temperature of rotary cement plant kiln is one of the most important variables by which clinker quality can be determined. It is very hard to develop a first principal model for prediction of kiln head temperature due to the presence of nonlinearity, time lag, and hysteresis in kiln operation. ten process variables were identified that affect kiln head temperature significant enough to be included in model development. Eighty percent of randomly selected plant data is used for training, optimization and testing and the balance of 20% is used for model validation. Support Vector Regression (SVR) and Artificial Neural Networks (ANN) are utilized to develop models for eight different sets of feature variables dictated by PCA and SVM to predict kiln head temperature. The parameters of these two models are optimized by Genetic Algorithm (GA) method and model predictions are compared. SVR-based model predictions with a minimum and maximum average absolute relative error of 0.742% and 1.413% outperformed ANN-based models.

回转水泥窑窑头温度是决定熟料质量的关键工艺参数之一。由于水泥窑运行过程中存在非线性、时滞与迟滞特性，难以建立用于预测窑头温度的第一性原理模型。研究筛选出对窑头温度影响显著的10项工艺变量用于模型构建。随机选取的现场实测数据中，80%用于模型训练、优化与测试，剩余20%用于模型验证。针对由主成分分析（Principal Component Analysis, PCA）与支持向量机（Support Vector Machine, SVM）生成的8组不同特征变量集，分别采用支持向量回归（Support Vector Regression, SVR）与人工神经网络（Artificial Neural Networks, ANN）构建窑头温度预测模型。通过遗传算法（Genetic Algorithm, GA）对两类模型的参数进行优化，并对比二者的预测性能。基于支持向量回归的预测模型表现更优，其平均绝对相对误差的最小值与最大值分别为0.742%与1.413%，优于人工神经网络模型。