Abstract: This paper proposes a hybrid model, which combines the seasonal time series autoregressive integrated moving average (SARIMA) and the generalized regression neural network (GRNN) models in order to forecast accurately urban short-term traffic flow. For comparison, the two component models, namely the SARIMA model and GRNN model, are used to forecast the short-term traffic flow. An empirical study shows the hybrid models have better forecast performance than the SARIMA model, but does not necessarily better than the GRNN mode. However, choosing proper input variables and output variables in the GRNN component of a hybrid plays an important role in improving the forecast ability of the model. The hybrid model constructed in this paper not only provides the best forecast performance but also has simple structure suitable for providing real-time and short-term traffic flow forecast.

Key words: short-term traffic flow forecasting, SARIMA model, GRNN model, hybrid model

摘要： 为了更精确地预测短期交通流，提出由季节自回归求和移动平均模型(SARIMA)和广义回归神经网络(GRNN)模型所构成的组合模型(SARIMA-GRNN模型)，该模型结合了时间序列模型和神经网络模型进行时间序列预测的优点。构造该组合模型的两个组成模型，即SARIMA模型和GRNN模型，也被用于预测研究以便于验证该组合模型在预测上的优势。实证研究结果表明，组合模型的预测精度高于SARIMA模型，但是却并不必然高于GRNN模型。然而，合理选择组合模型中神经网络部分的输入变量以及输出变量将显著地改善模型的预测精度，本文所构造的这个组合模型不仅具有很好的预测表现而且结构简单，非常适合城市道路短期交通流的实时预测。

关键词: 短期交通流预测, 季节自回归求和移动平均模型, 广义回归神经网络模型, 组合模型