摘要
针对钛合金棒材热连轧轧制力的精确预测问题,提出了一种基于加速收敛的粒子群(accelerate convergence particle swarm optimization,ACPSO)优化支持向量回归(support vector regression,SVR)的预测算法。该算法首先通过使粒子在每次速度迭代过程中偏离速度迭代一个小角度,在位置迭代过程中偏离迭代位置一小步,改善了粒子群算法的收敛性及收敛速度,再通过ACPSO算法实现对支持向量回归机的参数ε、c、γ的同时寻优,从而使ACPSO-SVR模型具有较高的预测精度和泛化能力。通过仿真实验和实际数据的比对,验证了方法的有效性。实验结果表明,ACPSO-SVR算法能够有效、快速地实现轧制力的精确预测,在预测速度和适应性方面,优于基于PSO-SVR(particle swarm optimization-support vector regression)的预测算法;在预测精度等方面,该算法优于BPNN(back propagation neural network)、SVR、PSO-SVR等算法,平均误差率从BP神经网络的±9%降到±4%以内。
Aiming at the accurate prediction issue of titanium alloy bar rolling force, a prediction algorithm based on support vector regression （SVR） optimized by accelerate convergence particle swarm optimization （ACPSO） is proposed. Firstly,this algorithm improves the convergence performance and convergence speed of particle swarm optimi- zation algorithm through making the particle speed deviate a small angle in each speed iterative process and making the particle position deviate a small step in each position iterative process. Secondly, ACPSO algorithm is applied to optimize the three parameters simultaneously, which makes the ACPSO-SVR model have good prediction accuracy and generalization capability. Simulation experiment and practical data comparison verify the validity of the method. Experiment results show that the ACPSO-SVR algorithm can effectively and quickly predict the rolling force; the ACPSO-SVR algorithm is superior to particle swarm optimization-support vector regression （PSO-SVR） algorithm in prediction speed and adaptability, and is better than back propagation neural network （BPNN）, SVR and PSO-SVR algorithms in prediction precision. The average error ratio decreases from ±9% achieved with BP neural network algorithm to less than ± 4% obtained with ACPSO-SVR algorithm.
出处
《仪器仪表学报》
EI
CAS
CSCD
北大核心
2012年第11期2579-2585,共7页
Chinese Journal of Scientific Instrument
基金
国家自然科学基金(61050006)资助项目
关键词
棒材热连轧
轧制力预测
支持向量回归
加速收敛粒子群优化
bar hot rolling
rolling force prediction
support vector regression （SVR）
accelerate convergence particle swarm optimization （ACPSO）
