极端学习机算法的改进及应用研究

摘要

极端学习机是一种新型的单隐藏层前馈神经网络模型，其输入权值和隐藏层阈值随机设置，其输出权值解析计算得到。因此，其运算速度是传统的 BP 神经网络的数千倍，而且具有良好的模型辨识能力。然而，极端学习机的输入权值和隐藏层阈值是随机设定的，可能不是使网络训练目标能达到全局最小值时的最优模型参数。针对此不足，本文采用最小二乘思想确定极端学习机的输入权值和隐藏层阈值。同时，将改进的极端学习机算法应用于电站锅炉的燃烧热效率建模，并与 BP、原始极端学习机、粒子群优化极端学习机和“教与学”优化极端学习机算法进行比较，证明了改进算法的有效性。%Extreme learning machine is a novel single hidden layer feed⁃forward neural network model,whose input weights and the bias of hidden nodes are generated randomly.And its output weights are computed analytically.Consequently,the extreme learning machine owns extremely fast speed and good identification ability,which is faster than conventional BP neural network thousands times.However,the stochastic input weights and the bias of the extreme learning machine are not the best model parameters possibly when the objective function gets the global minimum value.Therefore,the least square method is adopted to seek the appropriate pa⁃rameters of extreme learning machine.The improved extreme learning machine is applied to build the combustion thermal efficiency model of the plant boiler.Compared with other algorithms,such as BP,conventional extreme learning machine,particle swarm opti⁃mization extreme learning machine,teaching⁃learning⁃based optimization extreme learning machine,the result shows that the im⁃proved extreme learning machine is effective.