基于鲁棒相关向量机的转炉炼钢终点预报模型

针对传统相关向量机在训练过程中易受异常点影响的问题,提出了一种鲁棒相关向量机模型,并将其应用于转炉炼钢终点碳含量和温度的预报.通过为每一个训练样本设定独立的噪声方差系数,并使其在训练过程中随模型预测误差的增大而逐渐减小来降低异常点的影响,同时依据贝叶斯证据框架给出了模型超参数的迭代计算公式,进行参数的优化.使用标准测试数据和转炉炼钢实际生产数据进行仿真,结果表明本文模型具有较好的预报精度和鲁棒性.

Endpoint prediction model of basic oxygen furnace steelmaking based on robust relevance-vector-machines

To deal with the problem that the classical relevance vector machine is sensitive to outliers, we present a novel robust relevance vector machine. This machine is applied to predict the endpoint carbon content and temperature of the basic-oxygen-furnace(BOF) steelmaking. Each training sample is assumed to have its individual coefficient of noise variance. With the increase of the prediction error during training procedure, the coefficients of outliers gradually decrease, reducing the impact of outliers. In addition, the iterative formulas for the optimization of hyper-parameters are derived in the Bayesian evidence framework. Simulation results of benchmark test data and the BOF steelmaking data show that the proposed mode achieves high prediction accuracy and good robustness.