自适应BAS优化RBF神经网络的短时交通流预测

Short-term traffic flow prediction based on adaptive BAS optimized RBF neural network

摘要:

为提高短时交通流预测精度,针对传统径向基函数(radial basis function, RBF)神经网络短时交通流预测模型中心值固定、易受漂移数据干扰问题,提出自适应天牛须搜索算法(beetle antennae search algorithm, BAS)优化RBF神经网络的短时交通流预测模型。模型采用自适应步长提高BAS算法迭代速度和寻优能力,结合DBSCAN聚类确定RBF神经网络隐含层径向基函数网络中心,进而优化神经网络结构。通过路网真实交通流数据进行训练,选择常用于短时交通流预测的BP神经网络,RBF神经网络,广义RBF神经网络进行对比。结果表明:优化后的模型预测结果相较BP神经网络平均绝对误差降低了1.87%、平均绝对百分比误差降低了15.96%、均方根误差降低了3.24%,拟合度提高了3.96%；相较广义RBF神经网络平均绝对误差降低1.36%、平均绝对百分比误差降低了5.01%、均方根误差降低了2.19%,拟合度提高了2.5%。改进后的短时交通流预测模型能够为智能交通诱导提供可靠的预测值。

Abstract:

In order to improve the accuracy of short-term traffic flow prediction, an improved traffic flow predicting model was proposed by using the adaptive beetle antennae search (BAS) algorithm to optimize the radial basis function (RBF) neural network. This was done in consideration of the drawbacks of the traditional RBF neural network model for short-term traffic flow prediction, such as fixed center value and vulnerability to drift data interference. In this model, the adaptive step size was utilized to improve the iteration speed and optimization ability of BAS algorithm. The center of the RBF hidden layer was determined based on the DBSCAN cluster, and thus the neural network structure was optimized. Traffic flow datasets were collected from real road network for training, and the proposed model was compared with widely-used models, such as BP neural network, RBF neural network, and generalized RBF neural network. Results showed that in comparison with the BP neural network, the proposed method reduced the mean absolute error by 1.87, the mean absolute percentage error by 15.96, and the root mean square error by 3.24%; the fitting degree of the proposed method was improved by 3.96%. In comparison with the generalized RBF neural network, the proposed method reduced the mean absolute error by 1.36, the mean absolute percentage error by 5.01, and the root mean square error by 2.19%; the fitting degree of the proposed method was improved by 2.5%. The proposed short-term traffic flow prediction model can provide accurate predictions for intelligent traffic guidance.