多尺度时序特征融合的地铁短时客流预测方法

Short-term Metro Passenger Flows Prediction with Multi-Scale Temporal Feature Fusion

面向城市地铁进出站客流数据,本文提出了一种融合多尺度时序特征的地铁短时客流预测方法。首先,通过数据分析,构建地铁客流的长期周期性和短期波动性等多尺度时序特征融合机制。其次,提出地铁短时客流预测方法模型来有效提高客流预测精度,既利用了长期预测模型捕捉客流的长期周期性规律,又融合了短期预测模型获取客流的短期波动性趋势。最后,在2020年9月厦门地铁53个站点的客流数据上验证模型的有效性。实验结果表明,与LSTM、GRU等基线方法相比,本文提出的方法在地铁客流预测任务上具有预测精度更高、收敛速率更快等优势。

Based on the urban metro passenger flow data,a short-term metro passenger flows prediction method with multi-scale temporal feature fusion is proposed(shortly MAF).First,a fusion mechanism for metro passenger flows was constructed to capture the multi-scale temporal features,such as the long-term periodicity and the short-term volatility.Second,the MAF model was proposed to effectively improve the prediction accuracy by capturing the long-term periodicity patterns of passenger flows with long-term prediction method,and by obtaining the short-term volatility trend of passenger flows with the short-term prediction method.Finally,passenger flows data of 53 Xiamen metro stations in September 2020 was used to conduct the experiment and to testify the method effectiveness.The results show that,compared with LSTM,GRU and other models,our proposed MAF method outperforms baseline methods for metro passenger flows prediction tasks in terms of both the high prediction accuracy and the fast convergence rate.