改进贝叶斯动态线性模型在车流量预测中的应用

Application of modified Bayesian dynamic linear model on traffic flow prediction

贝叶斯动态线性模型(DLM)是一种状态空间模型,在时间序列分析和建模中有大量应用。传统DLM模型的参数确定是由人工分析时间序列的平稳性和季节、趋势、回归特征得到模型参数的,比较依赖专家经验。本文提出用期望最大化(EM)算法从交通流数据中学习出DLM的关键参数,然后用学习出的模型对下一时刻的车流量进行预测。实验表明该模型和方法应用在车流量的预测中取得了较好的效果。相对于自回归求和移动平均(ARIMA)模型和传统季节-趋势-回归组合DLM,所提方法在平均绝对误差和平均相对百分比误差两个指标上均有较大提高。

Bayesian dynamic linear model （DLM） is a kind of state space model which has a wide application on time series analysis and modeling. Traditionally, the parameters of DLM is determined by manually analyzing the stability and seasonal, trend and regression features of the time series, which is highly dependent on expert experience. In this paper, an expectation-maximization （EM） algorithm is proposed to learn the key parameters of DLM from traffic flow data, thereafter the model can be used to forecast the traffic flow at next moment. Experiments show that the model and learning method have a good performance on traffic flow prediction. It out-performs the traditional autoregressive integrated moving average （ARIMA） model and DLM composed by seasonal, trend and autoregressive components on mean absolute error and mean absolute percentage error.