基于时空残差网络的区域客流量预测方法

Regional Traffic Prediction Method Based on Spatiotemporal Residual Network

摘要:

针对区域客流量波动性强、复杂非线性的特征,易受到季节性影响,并且单一神经网络模型无法同时学习时间与空间相关性问题,通过对区域客流量影响因素分析,结合残差网络和全连接网络,提出了用于区域客流量预测的改进Quad-ResNet模型。Quad-ResNet模型融合了四个残差网络和一个全连接网络,该模型通过深层次的卷积学习空间相关性,结合四个残差网络学习时间邻近性、相似性、周期性、趋势性,使用全连接网络学习季节性影响。将Quad-ResNet模型与LSTM、CNN、ST-ResNet模型在同一数据集上进行区域客流量预测对比实验,实验结果表明,Quad-ResNet模型误差小于其他对比模型,而且在训练和预测的操作上明显比LSTM模型更简便,更适用于区域客流量预测。

Abstract:

For the characteristics of Regional traffic (Strong volatility, Complex nonlinearity, Susceptible to seasonal effects), as the single neural network model cannot learn temporal and spatial correlation problems simultaneously. By analyzing the influencing factors of regional tourist flow, combining residual networks with fully connected networks. The author proposes an improved Quad-ResNet model for regional tourist traffic forecasting. The Quad-ResNet model integrates 4 residual networks and a fully connected network, in learning spatial correlation through deep convolution, learning time proximity, similarity, periodicity and trend by combining 4 residual networks, also learning seasonal influencing factors by using a fully connected network. Comparing the Quad-ResNet model with the LSTM, CNN, and ST-ResNet models on the same data set for regional tourist traffic forecasting, the experimental results demonstrate that the deviation of Quad-ResNet model is smaller than other models. Moreover, it is obviously easier to train and forecast than the LSTM model, is more suitable for regional tourist traffic forecasting.