基于门控循环单元神经网络的公交到站时间预测

为提高用户公交出行积极性、方便管理部门合理调度公交班次,利用大数据分析公交浮动车辆历史GPS数据,考虑不同线路、公交站点地理位置、不同驾驶员、气象情况、时间分布等多因素的影响,建立了一种基于门控循环单元(gated recurrent unit, GRU)神经网络的公交到站时间预测模型。该模型结合5 000多万条原始数据,借助分布式Hadoop集群中的Spark弹性分布式数据集进行数据清理,并运用站点匹配算法进行源数据匹配、Lasso算法优化特征选项及去除干扰。实验仿真结果表明:改进的GRU模型R-square拟合度达到94.547%,并且算法效率较传统长短期记忆(long short-term memory,LSTM)神经网络提高了近14%,为进一步提高公交到站时间的预测精度与效率提供了参考。

Prediction of Bus Arrival Time Based on Gated Recurrent Unit Neural Networks

In order to increase the public transportation usage and the reasonability of the bus schedule by the management department, a novel prediction model of bus arrival time is proposed. This predicting model based on gated recurrent unit (GRU) neural network, analyzed the big data of historical GPS data about floating vehicle and considers the influence of different routes, bus station location, different drivers, weather conditions, time distribution and other factors. Furthermore, combining more than 50 million pieces of raw data, the model uses Spark elastic distributed data set in distributed Hadoop cluster to clean data and site matching algorithm to match source data, Lasso algorithm to optimize feature options and remove interference. The simulation results reveal that the R-square fitting degree of the improved GRU model is 94.547% and the prediction efficiency is nearly 14% higher than that of traditional long short-term (LSTM) model. It provides a reference for further improving the accuracy and efficiency of bus arrival time prediction.