基于深度神经网络的短时地铁客流预测

Short-Term Subway Passenger Flow Prediction Based on Deep Neural Network

摘要: 为更好推动城市轨道智慧化进程，利用复杂的深度神经网络对城市轨道交通客流进行短时精准预测，为客运交通运输规划提供科学的依据。针对传统的基于梯度下降法训练的长短期记忆网络交通流预测算法受网络初始值影响较大这一缺陷，提出了一种基于遗传算法改进的长短期记忆网络(LSTM)对郑州地铁一号线客流量进行短时预测。首先，搭建基于交通流预测的LSTM网络结构；然后，根据遗传算法通过选择、交叉、变异三种遗传操作不断迭代搜索得到LSTM最优的隐含层和全连接层层数以及对应的神经元个数；最后，利用训练好的LSTM神经网络模型对测试数据集进行预测。同时利用卷积神经网络从时间和空间两个维度进行预测，提取更高维度的站间客流特征，充分考虑了不同站之间客流量数据的关联性。实验结果表明，深度神经网络模型在短时地铁客流预测中具有较高的预测精度。

Abstract: In order to better promote the intelligentization of urban rail, a complex deep neural network is used to make short-term accurate predictions of urban rail transit passenger flow, and provide a scientific basis for passenger transportation planning. Aiming at the defect that the traditional Long Short-Term Memory network traffic flow prediction algorithm based on gradient descent training is greatly affected by the initial network value, a genetic algorithm-improved Long Short-Term Memory network (LSTM) is proposed for Zhengzhou Metro Line 1 Short-term forecast of passenger flow. First, building an LSTM network structure based on traffic flow prediction; then, according to the genetic algorithm, iteratively search through the three genetic operations of selection, crosso-ver, and mutation to obtain the optimal number of hidden and fully connected layers of the LSTM and the corresponding number of neurons; finally, use the trained LSTM neural network model to make predictions on the test data set. At the same time, Convolutional Neural Network is used to predict from the two dimensions of time and space, and the higher-dimensional characteristics of passenger flow between stations are extracted, which fully considers the correlation of passenger flow data between different stations. The experimental results show that the deep neural network model has high prediction accuracy in short-term subway passenger flow prediction.