变异PSO算法协同神经元网络在轧制力预报中的应用

Application of Adaptive Mutation PSO With Neural Networks to Rolling Force Prediction

为了避免BP神经元网络易陷入局部极值和基本粒子群(PSO)-神经元网络早熟收敛问题,采用一种自适应变异的粒子群优化算法训练神经元网络,根据轧制力的实测值和神经元网络的预报值确定粒子群算法的适应度函数,按照权重梯度方向进行变异操作,并首次将该方法应用到热连轧机组轧制力预报中。通过攀钢热轧板厂现场数据运算表明,该方法的预报误差平均值比传统数学模型低1.65%,比BP神经元网络低0.55%,收敛速度比BP神经元网络提高了约1/4,为进一步提高精轧机组轧制力预报精度提供了一种新的有效方法。

In order to avoid easily getting in local extremum like BP neural networks and premature convergence like PSO-neural networks, an improved particle swarm optimization algorithm with the adaptive mutation is used to train neural networks by means of determining sufficiency function of particle swarm optimization algorithm according to measured and predicted rolling force, after that variation operation along the direction of weight gradient starts, which is used for rolling force prediction of hot strip mills at first time. By off-line application for the data from hot strip mill of Panzhihua Steel, it has shown that the average prediction error of this method is 1.6 % lower than traditional model and 0.55% lower than BP neural networks, its convergent speed is improved by 1/4 than BP neural networks, so it provides a new valid method for improving prediction of rolling force of hot strip mill.