基于指数平滑法和残差网络的短时交通流预测方法

Short-term Traffic Flow Forecasting Model Research Based on Exponential Smoothing and Residual Network

针对短时交通流预测中短时波动特征提取预测困难、影响因素较多、与长时稳定特征混合难以分离的问题,提出1种基于三次指数平滑法(Triple Exponential Smoothing,TES)和深度残差网络(Deep Residual Networks,DRN)的组合预测模型.首先使用TES提取交通流数据中的趋势和周期特征,并得到剩余误差项;再使用DRN从剩余误差项中提取短时波动特征;最后融合上述2个模型的预测结果.基于北京巡游出租车和纽约共享自行车数据集进行验证.结果表明,在2种数据集上,只使用TES模型时均方根误差小于历史平均、自回归移动平均、季节性自回归移动平均和向量自回归模型;使用组合模型时,均方根误差小于循环神经、长短时间记忆、门控循环单元网络;相比于单独使用TES模型、DRN模型,组合模型在合适的网络深度下均方根误差降低4%以上.结果证明,组合模型可提高短时交通流预测准确率,且预测效果优于传统时间序列预测模型和上述神经网络模型.同时在需要大幅减少模型计算成本时,仅单独使用TES模型也能达到较高精度.

The short-term fluctuation features are difficultly extracted and predicted;influenced by many factors;are difficultly separated from long-term stable features.Aiming at the problems mentioned,a combined forecasting model based on Triple Exponential Smoothing(TES)and Deep Residual Network(DRN)is introduced.First,TES is used to extract the trend and seasonal components in traffic flow data,and obtain the residuals.Then,DRN is used to extract short-term fluctuation features from the mentioned residuals.Last,the forecasting results of these two models are merged.The model evaluation was implemented based on the datasets of Beijing cruising taxicabs and New York sharing bikes.The evaluation results showed that the root mean squares error(RMSE)of single TES model was less than that of Historical Average(HA),Autoregressive Integrated Moving Average(ARIMA),Seasonal ARIMA(SARIMA)and Vector Auto-Regressive(VAR)in the both datasets mentioned with the different size of samples;the RMSE of the combined model was less than that of Recurrent Neural Network(RNN),Long-short Term Memory(LSTM)and Gated Recurrent Unit Network(GRU).Compared with single TES and DRN model,the RMSE of combined model decreased by at less 4%with the optimal network depth.The evaluation results have proved that the combined model can improve the short-term traffic flow forecasting accuracy,and its performance is better than other general time series forecast models and mentioned neutral network models.The single TES model can realize high forecasting accuracy when the computational cost of model is need to be remarkably decreased.