A real-time data forecasting method for dynamic system is proposed to improve the data accuracy of power station,which applies the Elman neural network evolved by improved PSO(Particle Swarm Optimization) algorithm. In the improved PSO algorithm, the mutation operation for the inactive particles is carried out according to two factors:the premature convergence degree of the swarm and the current optimal solution,which enhances its ability to break away from local optimum. The improved PSO algorithm is used to evolve the network weight and self-feedback coefficient of Elman neural network to avoid the defects of the gradient descent algorithm:the slow convergence of weight learning and the premature result. Case study for the main steam flow rate of 300 MW power plant shows that, it obtains the system dynamics properly with excellent denoise ability and robustness to abnormal data. Compared with standard Elman neural network,it has better forecasting accuracy and generalization ability.%为了提高电站实时数据的准确性,提出了一种利用改进粒子群算法进化Elman神经网络的动态系统实时数据预测方法.改进粒子群算法中,根据群体早熟收敛程度和当前最优解的大小对部分不活跃粒子进行变异,增强了算法跳出局部最优解的能力.利用改进的粒子群算法训练Elman神经网络权值和自反馈增益因子,有效地解决了梯度下降法训练网络权值收敛速度慢、易陷入局部极值的缺点.以某300 MW机组的主蒸汽流量为具体对象,给出了该方法的算例,结果表明该方法能正确获取系统动态特性,具有较强的降噪能力,对异常数据具有鲁棒性.与标准Elman神经网络进行比较,该方法具有较好的预测精度和泛化能力.
展开▼
