574 工程科学学报,第43卷,第4期 15o0 1500(b 1450 Measured values 1450 values -SVR predictive values Test set sample numbers:1-100 Test set sample numbers:1-100 1500 1500 1450 Measu ed values 1450 Measured values -SVR predictive values Test set sample numbers:101-200 ELM predictive values Test set sample numbers:101-200 1500 1500 1450 Measured values 1450 -SVR predictive values Test set sample numbers:201-300 Test sct sample numbers201-300 1500 8 8 1500 见点安d口 1450 Measured values 1450 Measured values -SVR predictive values Test set sample numbers:301-400 ELM predictive values Test set sample numbers:301-400 20 40 60 80 100 0 20 40 60 80 100 Sample number Sample number 图3铁水温度测量值与预测值比对.(a)基于SVR算法:(b)基于ELM算法 Fig.3 Comparison of measured and predictive values of hot metal temperature:(a)prediction value based on support vector regression(SVR); (b)prediction value based on extreme learning machine(ELM). 图5为SVR和ELM对铁水温度预测结果的 (a) 1510 用日 百分比误差散点分布与统计图.在0.5%范围内 的点个数除上所有点的个数,即为分布概率经统 1490 计计算,SVR和ELM的预测百分比误差在0.5% 1470 以内的分布概率分别为81.0%和78.0%,在±1.0%以 SVR predictive value and 内的分布概率分别为99.0%和983%,说明百分比 measured value comparison 1450 误差主要集中在±0.5%,基本不超过±1.0%.对比两 1450 1470 1490 1510 Measured hot metal temperature/C 种算法预测误差散点分布图,SVR算法更好地控 制了误差偏离严重预测点的数量,因而百分比误 (b) 1510 差点波动较低 0 100 200 300 4000 20 4060 1490 0 1.0. ■SVR percentage error SVR error 0 0.5 1470 g 0 -0.5 -1.0 。g (a) 14 1450 1470 1490 1510 Measured hot metal temperature/C 1.0 ■ELM percentage erro ELMerro 0.5 0 0.08 -0.5 -1.0 0.06 (b) 100 200 300 4000 204060 0.04 图5铁水温度预测的百分比误差散点分布统计图.(a)SVR:(b)ELM 0.02 Fig.5 Scatter distribution statistics of percentage error in hot metal temperature prediction:(a)SVR:(b)ELM -20 -10 0 10 20 对SVR和ELM算法预测结果进行综合分析 Hot metal temperature error/C 算法预测准确度一般应从多方面予以定量化衡量 图4铁水温度预测值与测量值偏差.()基于SVR的铁温预测值与 与表征],本文选取平均绝对误差(Mean absolute 测量值偏差:(b)基于ELM的铁温预测值与测量值偏差:(c)基于 error,,MAE)用以表征模型预测值与测量值之间整 SVR与ELM预测铁温误差概率密度分布函数 体偏差的绝对值;选取平均绝对百分比误差(Mean Fig.4 Deviation of predictive value of hot metal temperature from the absolute percentage error,MAPE)用以表征模型预测 measured value:(a)based on SVR;(b)based on ELM;(c)the probability density distribution function of hot metal temperature error based on SVR 值与测量值之间整体偏差的相对值;选取均方根 and ELM 误差(RMSE,Root mean square error)用以表征模型图 5 为 SVR 和 ELM 对铁水温度预测结果的 百分比误差散点分布与统计图. 在±0.5% 范围内 的点个数除上所有点的个数,即为分布概率. 经统 计计算,SVR 和 ELM 的预测百分比误差在±0.5% 以内的分布概率分别为 81.0% 和 78.0%,在±1.0% 以 内的分布概率分别为 99.0% 和 98.3%,说明百分比 误差主要集中在±0.5%,基本不超过±1.0%. 对比两 种算法预测误差散点分布图,SVR 算法更好地控 制了误差偏离严重预测点的数量,因而百分比误 差点波动较低. 1.0 0 100 200 300 SVR percentage error SVR error point distribution ELM error point distribution ELM percentage error 400 0 100 200 300 400 0 20 40 60 0 20 40 60 0.5 −0.5 −1.0 0 Percentage error/ % 1.0 0.5 −0.5 −1.0 0 Percentage error/ % (a) (b) 图 5 铁水温度预测的百分比误差散点分布统计图. (a)SVR;(b)ELM Fig.5 Scatter distribution statistics of percentage error in hot metal temperature prediction: (a) SVR; (b) ELM 对 SVR 和 ELM 算法预测结果进行综合分析. 算法预测准确度一般应从多方面予以定量化衡量 与表征[25] ,本文选取平均绝对误差(Mean absolute error,MAE)用以表征模型预测值与测量值之间整 体偏差的绝对值;选取平均绝对百分比误差(Mean absolute percentage error,MAPE)用以表征模型预测 值与测量值之间整体偏差的相对值;选取均方根 误差(RMSE,Root mean square error)用以表征模型 Test set sample numbers: 1-100 Test set sample numbers: 1-100 Test set sample numbers: 101-200 Test set sample numbers: 201-300 Test set sample numbers: 301-400 Test set sample numbers: 101-200 Test set sample numbers: 201-300 Test set sample numbers: 301-400 Sample number 1500 (a) (b) 1450 1500 1450 1500 1450 1500 1450 Temperature/ ℃ 1500 1450 1500 1450 1500 1450 1500 1450 Temperature/ ℃ 0 20 40 60 80 100 Sample number 0 20 40 60 80 100 Measured values SVR predictive values Measured values SVR predictive values Measured values SVR predictive values Measured values SVR predictive values Measured values ELM predictive values Measured values ELM predictive values Measured values ELM predictive values Measured values ELM predictive values 图 3 铁水温度测量值与预测值比对. (a)基于 SVR 算法;(b)基于 ELM 算法 Fig.3 Comparison of measured and predictive values of hot metal temperature: (a) prediction value based on support vector regression (SVR); (b) prediction value based on extreme learning machine (ELM). Predictive hot metal temperature/ ℃ Measured hot metal temperature/℃ 1510 1510 1490 1490 1470 1470 1450 1450 (a) SVR predictive value and measured value comparison Predictive hot metal temperature/ ℃ Measured hot metal temperature/℃ 1510 1510 1490 1490 1470 1470 1450 1450 (b) ELM predictive value and measured value comparison Probability density of prediction error Hot metal temperature error/℃ 0.08 0.06 0.04 0 0.02 (c) Probability density of SVR prediction error Probability density of ELM prediction error −20 −10 0 10 20 图 4 铁水温度预测值与测量值偏差. (a)基于 SVR 的铁温预测值与 测量值偏差;(b)基于 ELM 的铁温预测值与测量值偏差;(c)基于 SVR 与 ELM 预测铁温误差概率密度分布函数 Fig.4 Deviation of predictive value of hot metal temperature from the measured value: (a) based on SVR; (b) based on ELM; (c) the probability density distribution function of hot metal temperature error based on SVR and ELM · 574 · 工程科学学报,第 43 卷,第 4 期