【作者】 刘军；

【导师】 袁平；

【作者基本信息】 东北大学 ， 控制理论与控制工程， 2013， 硕士

【摘要】 电弧炉炼钢是现代大规模炼钢的方法之一,同时也是生产特殊钢和高合金钢的主要方法。以废钢为主要原料的电弧炉炼钢在世界各国稳步地发展。电弧炉炼钢是在高温下进行的,但是由于测量手段和成本的限制,使得对钢水温度的实时测量变得尤为困难,因此基于软测量技术的钢水终点温度预报显得相当重要。本文首先介绍了电弧炉炼钢的工艺过程,详细分析了冶炼过程中的能量收支情况,并在此基础上总结了影响钢水终点温度的主要因素。鉴于案例推理技术具有知识获取容易、推理简单、能够自学习的优点,本文将案例推理技术引入到对钢水终点温度预报的问题中,建立了基于案例推理的钢水终点温度预报模型。但匹配案例的权重值的线性确定方式和不加补偿的重用导致了预报模型的预报命中率和预报精度偏低。在案例推理技术中,匹配案例权值系数的确定方式将对当前案例的解产生直接的影响。在原有模型中,权值系数的确定方式会削弱相似度较大案例对当前案例解的参考价值。针对这个问题,本文提出了一种改进的权值系数的确定方式,提高了相似度较大案例对当前案例解的权重值。仿真结果表明,基于改进的案例推理的终点温度预报模型,在预报命中率和精度方面都有较大提高。由于当前案例与匹配案例之间总是存在差异的,不加修正的重用匹配案例必然会产生较大的误差,因此本文首先采用线性表达的增量模型对匹配案例进行补偿,然后再对补偿后的案例进行重用。但考虑到电弧炉炼钢过程具有严重的非线性,本文进而采用非线性的BP神经网络取代原有的增量模型对匹配案例进行补偿。由于BP神经网络本身可能存在的收敛速度慢、容易陷入局部极小的缺点,本文采用粒子群算法对其进行了优化。本文最后对基于补偿的案例推理的钢水终点温度预报模型进行了仿真,结果表明改进后的预报模型具有更高的预报命中率和精度。更多还原

【Abstract】 Electric arc furnace (EAF) steelmaking is one of the modern large-scale steelmaking methods and also the main way to produce special steel and high alloy steel. EAF steelmaking with the scrap steel as its main material, gets a substantial development all over the world. EAF steelmaking is carried out at a high temperature. The real-time measurement of molted steel becomes hard for the limits of the measuring means and costs, so prediction of molted steel end-point temperature based on soft-sensing technique is very important.In this paper, the craft process of EAF steelmaking is firstly introduced, and then the energy budget during the process is analyzed in detail. Based on the analysis, the main factors influencing the end-point temperature of molted steel are gained. And then, the prediction model of end-point temperature of molted steel in EAF based on case-based reasoning (CBR) is built for the advantages of CBR, such as easiness to acquire knowledge, simple reasoning and the ability of self-learning. But, the method to ascertain the weight coefficients of the cases matched in a linear way and the reuse without any compensation of them in CBR cause the hit rate and precision of the prediction model lower.In CBR, the way to ascertain the weight coefficients of the cases matched has a direct influence on the solution of the current case. In the original model, the way to ascertain the weight coefficients would weaken the reference value of the cases with higher similarities. So a modified way to ascertain the weight coefficients is proposed to raise the weights of the cases with higher similarities. The simulation result indicates the modified model has an advantage on the hit rate and precision over the original model.The reuse without any revision of the cases matched would cause a large error for the differences between the current case and them. To solve the problem, an incremental model with a linear expression is proposed to compensate the solutions of the cases matched. Then the cases compensated are reused. But the process of EAF steelmaking is one with high nonlinearity. A nonlinear compensation model based on back-propagation neural network (BPNN) is proposed to replace the original incremental model. Further more, an algorithm of particle swarm optimization (PSO) is used to overcome the shortcomings of BPNN, such as a low rate of convergence and easiness to fall into local minimums.In the final part of this paper, the prediction model of end-point temperature of molted steel in EAF based on CBR compensated by PSO-BPNN is simulated and the result indicated the model has a higher hit rate and better precision.更多还原