摘要： 加油站是重要的能源供给单位，对加油站站点的下一时段客流量进行精准预测，可为相关资源的调度与分配提供决策支撑。针对加油站级客流量预测问题，结合加油站客流数据的时空特征，提出一种基于注意力机制的时空网络模型。以路网结构建模的站级客流数据为输入，结合卷积神经网络、长短期记忆网络与注意力机制，解决站点间的空间依赖、短期与长期时序依赖以及长期时序依赖中的时间漂移问题，精准预测下一时段的站级客流量。在真实数据集上的实验结果表明，与历史平均模型、长短期记忆网络模型和双向长短期记忆网络模型等基线模型相比，该模型在均方误差（RMSE）、平均绝对误差与平均绝对百分比误差上均有所提升，其中RMSE提升22.89%。

关键词: 时空数据, 客流量预测, 注意力机制, 卷积神经网络, 长短期记忆网络

Abstract: Gas station is an important energy supply unit, the accurate prediction of the foot traffic of each gas station in the next period can provide important support for the scheduling and allocation of related resources. Aiming at the problem of the gas station-level foot traffic prediction, combined with the spatial-temporal characteristics of gas station foot traffic data, this paper proposes an Attention Mechanism-based Spatial Temporal Network (AMSTN) model. Taking the station-level foot traffic data modelled in road network structure as input, the model integrates Convolution Neural Network (CNN), Long Short-Term Memory (LSTM) network and attention mechanism to deal with the spatial dependence, short-term and long-term temporal dependence, and time drifting in long-term temporal dependence between stations. On this basis, the accurate prediction of the station-level foot traffic in the next period is realized. The experimental results on real data sets show that compared with the Historical Average (HA) model, LSTM network model and bidirectional LSTM network model, the proposed model improves Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE), and its RMSE is increased by 22.89%.

Key words: spatial-temporal data, foot traffic prediction, attention mechanism, Convolution Neural Network(CNN), Long Short-Term Memory (LSTM) network

中图分类号: 

• TP391.4