基于情境感知注意机制的出租车需求深度学习预测模型

DEEP LEARNING PREDICTION MODEL OF TAXI DEMAND BASED ON CONTEXT AWARE ATTENTION MECHANISM

针对出租车需求预测方法中由于未充分挖掘数据导致的预测精度较低问题,提出一种基于情境感知注意机制的递归卷积神经网络出租车需求预测方法。根据城市路网划分出一组细粒度区域,得到出租车需求的三个时空依赖性预测结果。设计由局部卷积层和门控神经单元组成的即时时空模块,提取相邻区域和函数相似区域之间的空间相关性,以及连续时隙中的时间相关性。基于循环神经网络提出一种新的情境感知注意机制,将每个区域的三个预测合并在一起,将情境因素输入到完全连接层中来学习分配给每个预测的权重。以纽约市和成都市的实际数据集为基础进行综合实验,结果表明所提出方法有较高的预测精度以及运行时间优势。

In order to solve the problem of low prediction accuracy caused by insufficient data mining in taxi demand forecasting method,a recursive convolution neural network taxi demand forecasting method based on context aware attention mechanism is proposed.A group of fine-grained regions were divided according to the urban road network,and three spatiotemporal dependence prediction results of taxi demand were obtained.Furthermore,a real-time spatiotemporal module composed of local convolution layer and gate current unit was designed to extract the spatial correlation between adjacent regions and function similar regions,as well as the temporal correlation in continuous time slots.A new context aware attention mechanism based on recurrent neural network was proposed.The three predictions in each region were combined together,and the context factors were input into the fully connected layer to learn the weight assigned to each prediction.Comprehensive experiments were carried out based on the actual data sets of New York City and Chengdu City.The results show that the proposed method has higher prediction accuracy and running time advantages.