Abstract: In order to establish an accurate and effective traffic accident prediction model and improve the level of expressway traffic safety, 65,119 cases of traffic accidents on 11 expressways in Chongqing from 2011 to 2016 were used as the research object, and two total indicators such as “number of accidents” and “deaths toll” were selected to describe the monthly distribution pattern of highway traffic accidents in the time dimension.The autoregressive differential moving average (ARIMA) model was used to capture the linear time series characteristics in the time series data, and the long and short-term memory neural network (LSTM) model was used to fit the nonlinear time series characteristics in the prediction residual series.A combined prediction model for highway traffic accidents based on ARIMA and LSTM has been established, with RMSE and MAPE as evaluation indicators for the model.The results show that the prediction accuracy of each index of the ARIMA-LSTM combined prediction model is better than that of the single ARIMA model.Among them, the “death toll” combined model has a significant improvement effect, and the root mean square error (RMSE) and mean absolute percentage error (MAPE) are respectively improved by 55.83% and 54.80%, compared with ARIMA model.The RMSE and MAPE of the “number of accidents” combined model are respectively improved by 23.15% and 23.29%, compared with ARIMA model.

Key words: traffic engineering; traffic accident prediction; ARIMA-LSTM; combination model; expressway; time series

摘要： 为建立准确有效的交通事故预测模型，提升高速公路交通安全水平，以重庆市11条高速公路2011—2016年共计65 119起交通事故为基础，选取“事故数量”和“死亡人数”2项总量指标，描述统计高速公路交通事故在时间维度上的月分布规律。通过自回归差分移动平均（ARIMA）模型捕捉时间序列数据中的线性时序特征，使用长短时记忆神经网络（LSTM）模型拟合预测残差序列中的非线性时序特征，建立了基于ARIMA和LSTM的高速公路交通事故组合预测模型，并以均方根误差(RMSE)、平均绝对百分比误差（MAPE）值作为模型的评估指标。结果表明：ARIMA-LSTM组合预测模型各项指标的预测精度均优于单一的ARIMA模型，其中“死亡人数”组合模型改善效果显著，其RMSE与MAPE值相较于ARIMA模型分别改善了55.83%和54.80%；“事故数量”组合模型的RMSE和MAPE相较于ARIMA模型改善了23.15%、23.29%。

关键词: 交通工程；交通事故预测；ARIMA-LSTM；组合模型；高速公路；时间序列