[关键词]
[摘要]
采用燃煤机组脱硫系统原始非高斯数据建立的单一出口SO2质量浓度预测模型精度较低,泛化能力较差,针对该问题提出一种基于高斯混合模型-支持向量回归(Gaussian Mixture Model-Least Squares Support Vector Regression,GMM-LSSVR)的建模方法。采用高斯混合模型(GMM)将训练集数据聚类为多个高斯数据集,对每个对应的子集建立独立的最小二乘支持向量回归机(LSSVR)训练模型。在此基础上,估计测试集数据属于每个种群的概率并对测试集进行聚类,将每个子集输入到对应的LSSVR模型中完成预测。现场数据实验表明:采用GMM聚类后每个子集的概率密度不规则波动幅度减小,数据高斯性增强;GMM-LSSVR建模方法测试集均方根误差(RMSE)、 平均绝对误差(MAE)和可决定系数(R2)较单模型LSSVR方法有较大改善,具有更好的预测精度和泛化性能。
[Key word]
[Abstract]
The single outlet SO2 concentration prediction model established by using the original nonGaussian data of the desulfurization system of the coal-fired unit has low precision and poor generalization ability. To solve this problem, this paper proposes a modeling method based on Gaussian mixture model least squares support vector regression (GMM-LSSVR). Gaussian mixture model (GMM) is used to cluster the training set data into multiple Gaussian data sets. An independent LSSVR training model is established for each corresponding subset. On this basis,the probability that the test set data belongs to each population is estimated to cluster the test set, and the data of each subset is input into the corresponding LSSVR model to complete the prediction. The field data experiment shows that the irregular fluctuation of probability density of each subset is reduced and the Gaussian property of data is enhanced after GMM clustering; GMM-LSSVR modeling method has greatly improved the root mean square error (RMSE),mean absolute error (MAE) and determinable coefficient (R2) of the test set compared with the single model LSSVR method, which has better prediction accuracy and generalization performance.
[中图分类号]
TM621
[基金项目]