基于改进可变学习速率BP算法的短时交通流预测

Variable learning rate based on improved BP algorithm of short-term traffic flow forecasting

为了解决可变学习速率的BP神经网络(VLBP)在训练时容易陷入局部极小的问题,在VLBP的算法规则中引入模拟退火中的metropolis接受准则,使得在均方误差增量超过设定的界限值时,权值更新不总是被取消,而是以一定的概率被接受,构造了一种容易跳出局部极小的VLBP神经网络(MVLBP)。运用MVLBP算法对短时交通流进行预测,仿真结果表明,MVLBP神经网络训练收敛速度更快,且有较好的预测精度。

In order to solve the local minimum problems of variable learning rate BP neural network( VLBP) in training,the metropolis acceptance criteria in the SA algorithm is introduced to VLBP neural network. Then the weight update is not always cancelled,but accepted with a certain probability when the increment of the mean square error exceeds the preset value,a modified VLBP network which can easily hop from the local minimum is constructed( MVLBP). The MVLBP neural network is applied to predict short-term traffic flow. The simulation results show that this method improves the convergence speed and forecasting precision