为预测煤矸石代黏土煅烧水泥的28 d 抗压强度性能，根据生产水泥的物检分析数据，将GM(1，N)预测技术和径向基函数（RBF）神经网络技术相结合，提出了基于GM-RBF 神经网络组合模型水泥强度预测的新方法。该组合模型首先利用试验产品的典型物检数据建立GM(1，N)网络，对数据进行预处理。然后将输入样本数据进行一次累加生成操作，并进行归一化，设置GMRBF神经网络组合模型预测精度和散步常数。经处理后的输入样本作为RBF 神经网络输入向量，相应的实测28 d 抗压强度作为模型的输出期望值开展训练，比较预测数据与实测数据，并进行调整，最终得到符合精度要求的GM-RBF 神经网络组合模型。该组合模型一方面避免GM(1，N)模型的理论误差，利用累加生成运算和样本数据的预处理，减少了由于训练样本随机性对建模精度产生的影响；另一方面由于具有自适应、自组织和速度快等特点，能快速预测水泥远期强度情况。仿真试验表明，该模型预测精度优于单个GM（1，N）模型或RBF 神经网络模型，具有较好的拟合性，适用于对水泥强度的预测，可以为煤矸石代黏土煅烧水泥的质量分析提供有效参考。
In order to predict the 28 day compressive strength of coal gangues as clay for cement, a prediction method of grey model and radial basis function (GM-RBF) neural network combination model is presented according to the data of cement physics test analysis. The method makes used of the advantage of both GM and RBF neural network. Firstly, the combination model built up GM(1, N) network based on its test analysis data, and the data were preprocessed. One accumulated generation operation (1-AGO) and normalization were carried out. Predicted precision and scatter constant were set up to the combination model. Then, these processed samples served as the input vectors for RBF neural network, the measurement data of 28 day compressive strength served as output expectation value for model. Comparisons were carried out between prediction data and measurement data, then the data were adjusted logically. Finally, the GM-RBF neural network combination model is fit for precision requirement. AGO and pretreatment were used for data processing which can reduce randomness of training samples. It also shows that the method is self-adaptive, self-organized and fast. The model can not only avoid the theoretical error of GM(1,N), but also predict the further period compressive strength. The results show that it is better than GM(1,N) model or RBF neural network model. The combination model owns a fine agreement and adapts to predict cement strength. It can provide efficient reference of quality analysis for coal gangues as clay for cement.
[1] Antoni M, Rossen J, Martirena F, et al. Cement substitution by a combi-nation of metakaolin and limestone[J]. Cement and Concrete Research, 2012, 42(12): 1579-1589.
[2] Choi Y W, Kim Y J, Choi O, et al. Utilization of tailings from tungsten mine waste as a substitution material for cement[J]. Construction and Building Materials, 2009, 23(7): 2481-2486.
[3] Qiu G H, Luo Z Y, Shi Z L, et al. Utilization of coal gangue and copper tailings as clay for cement clinker calcinations[J]. Journal of Wuhan Uni-versity of Technology: Materials Science Edition, 2011, 26(6): 1205-1210.
[4] 李晓东, 杨波, 董吉文. 基于遗传算法的水泥强度预测[J]. 计算机工程 与应用, 2004, 40(12): 205-207. Li Xiaodong, Yang Bo, Dong Jiwen. Predicting of cement strength based on genetic algorithm[J]. Computer English and Applicaions, 2004, 40 (12): 205-207.
[5] 刘思峰, 党耀国, 方志耕, 等. 灰色系统理论及其应用[M]. 北京: 科学出 版社. 2010: 169-171. Liu Sifeng, Dang Yaoguo, Fang Zhigeng, et al. The theory of grey system and application[M]. Beijing: Science Press, 2010: 169-171.
[6] 鲍雅萍, 马金元, 宋强. 基于灰色神经网络的烧结矿碱度组合预测[J]. 控制理论与应用, 2008, 25(4): 791-793. Bao Yaping, Ma Jinyuan, Song Qiang. Combination forecasting of sin-tered ore alkalinity based on grey neural network[J]. Control Theory & Application, 2008, 25(4): 791-793.
[7] Alexandridis A, Triantis D, Stavrakas I, et al. A neural network approach for compressive strength prediction in cement-based materials through the study of pressure-stimulated electrical signals[J]. Construction and Building Materials, 2012(30): 294-300.
[8] 张筠莉, 杨祯山. 现代医院门诊量的灰色RBF神经网络预测[J]. 计算机 工程与应用, 2010, 46(29): 225-228. Zhang Yun1i, Yang Zhenshan. Grey RBF neural network based forecast-ing of outpatient capacity in modern hospita1[J]. Computer Engineering and Applications, 2010, 46(29): 225-228.
[9] 韩涛. 矿渣粉粒度分布特征及其对水泥强度的影响[D]. 西安: 西安建 筑科技大学, 2004. Han Tao. The distribute characteristic of slag powder and the influence on strength of cement[D]. Xi'an: Xi'an University of Architecture and Technology, 2004.
[10] 景晓璐, 张小栋, 耿加民. 基于RBF神经网络模型和SVM模型的压 力传感器温度补偿方法[J]. 科学技术与工程, 2012, 12(26): 6666-6670. Jing Xiaolu, Zhang Xiaodong, Geng Jiamin. RBF and SVM based temperature compensation algorithms for pressure sensor[J]. Science Technology and Engineering, 2012, 12(26): 6666-6670.
[11] 闻新, 周露, 王丹力, 等. MATLAB神经网络应用设计[M]. 北京: 科 学出版社, 2001: 245-246. Wen Xin, Zhou Lu, Wang Danli, et al. Application design of neural networks on MATLAB[M]. Beijing: Science Press, 2001: 245-246.
[12] 郭一军, 段杏林. 基于RBF神经网络水泥强度预测模型的研究[J]. 工业仪表与自动化装置, 2009(2): 99-101. Guo Yijun, Duan Xinglin. Forecast of cement strength based on RBF neural networks[J]. Industrial Instrumentation & Automation, 2009 (2): 99-101.
[13] 杨祯山, 邵诚. 基于灰色RBF神经网络模型的电梯层间交通分布预测[J]. 信息与控制, 2008, 37(6): 690-696. Yang Zhenshan, Shao Cheng. Forecasting of elevator interflow traffic distribution based on grey RBF neural network[J]. Information and Control, 2008, 37(6): 690-696.
[14] 李晓东. 基于数字图像处理的水泥强度预测模型研究[D]. 济南: 济南 大学, 2004: 35-36. Li Xiaodong. Prediction mode research of cement strength based on digital image processing[D]. Jinan: University of Jinan, 2004: 35-36.
[15] 杨丹. RBF神经网络预测水泥水化热研究[J]. 国防交通工程与技术, 2011, 9(3): 31-37. Yang Dan. The prediction of the hydration heat of cement with the RBF neural network[J]. Traffic Engineering and Technology for National Defence, 2011, 9(3): 31-37.
