基于自适应并联结构神经网络的交通流量预测

A Neural Network with Adaptive Parallel Structure and Its Application to Traffic Flow Prediction

摘要:

由于现有优化算法在全局优化方面的局限性，导致神经网络需要多次训练才能获得满意的结果。为了解决神经网络训练中的一致性问题，文章提出了一种自适应并联结构神经网络（Adaptive Parallel Structure Neural Network, APSNN）。APSNN由多个神经网络单元并联组成，在训练过程中，采用常规优化算法对各神经网络单元进行训练。神经网络单元的训练样本由上一级神经网络单元的训练残差构成，通过训练残差在各神经网络单元中的单向传递，实现训练残差的逐级减少。神经网络根据训练残差，决定是否进行神经网络单元级联和结构扩张，从而保证训练结果的一致性。文章对5种非线性函数进行了神经网络逼近测试。与BP神经网络相比较，APSNN在50次不同初始条件下，训练精度十分稳定，具有很好的一致性。为了实现对交通流量预测，文章将APSNN与BP神经网络和小波神经网络进行了对比研究，结果表明：APSNN的预测总体标准差均小于BP神经网络和小波神经网络，交通流量的预测偏差较BP神经网络和小波神经网络分别降低2.7%和9.7%。

Abstract:

Due to the limitations of existing optimization algorithms in global optimization, neural networks require multiple trainings to obtain satisfactory results. In order to solve the consistency problem of neural network training, this paper proposes an adaptive parallel structure neural network (APSNN). APSNN consists of multiple neural network units in parallel, and each neural network unit is trained using conventional optimization algorithms during the training process. The training samples of the neural network unit are composed of the training residuals of the previous neural network unit. Through the one-way transmission of the training residuals in each neural network unit, the training residuals are gradually reduced. According to the training residuals, the neural network decides whether to cascade the neural network units and expand the structure, so as to ensure the consistency of the training results. The neural network approximation test is carried out on five nonlinear functions in this paper. Compared with BP neural network, APSNN has very stable training accuracy and good consistency under 50 different initial conditions. In order to predict traffic flow, this paper compares APSNN with BP neural network and wavelet neural network. The results show that the overall standard deviation of APSNN prediction is smaller than that of BP neural network and wavelet neural network, and the prediction deviation of traffic flow is 2.7% and 9.7% lower than that of BP neural network and wavelet neural network, respectively.