基于LSTM与时空结合的共享单车站点需求量预测

Spatial-temporal LSTM for Demand Prediction of Shared Bike Stations

由于不同站点的车辆需求存在着时空分布不均衡现象,特别是早晚高峰期人流量大和车辆使用量波动性大的站点,以往的基于历史系统出行数据、时间和天气数据进行预测的模型预测性能下降.为此,本文提出了一种基于LSTM与时空结合的方法对不同类型站点的车辆需求量进行细粒度预测,并与HA、BP神经网络和GBDT模型的预测结果进行比较.实验结果表明,在加入时空结合特征后,LSTM模型针对借还车辆较大和波动性大的站点预测误差小,能够很好地在站点需求量变化大时跟踪其变化趋势.

Due to the uneven distribution of bike demands at different stations,especially at the stations with high peak traffic volume in the morning and evening and with high volatility of bike usage,the traditional demand prediction models,which are based on historical travel data and time and weather data to make predictions,do not work well.Therefore,this paper proposes a method based on the combination of LSTM and time and space integration to predict the bike demands of different types of stations in a fine-grained manner,and compares the prediction results with those from the HA,BP neural networks and the GBDT model.The experimental results show that the LSTM model in combination with the spatiotemporal integration feature can better track the change trend when the demand of the station changes greatly,as the prediction error is smaller for the stations with large numbers of borrowed bikes and high volatility of usage.