Deep Cascade Convolutional Neural Network For Spatiotemporal Predictive Learning

With the development of Io T network and the popularization of smart devices like mobile phones,spatiotemporal data containing both temporal and spatial information is continuously collected all the time.Due to its various sources and large amount,spatiotemporal data has gradually become an important data type and is widely used in many fields.In the application of spatiotemporal data,people usually need to predict spatiotemporal data for a period of time in the future with historical sequences,indicating that spatiotemporal predictive learning is an important direction among spatiotemporal data analysis and applications.The core of this problem is to simultaneously capture and combine temporal and spatial information when making predictions.However,existing methods for modeling temporal dynamics rely on recurrent neural networks,whose spatial abstraction ability is greatly limited,while for spatial correlations modeling,current approaches mainly adopt convolutional neural networks,where temporal dynamics cannot be modeled effectively.To deal with the dilemma of simultaneous modeling of temporal dynamics and spatial correlations,this paper proposes a deep cascade convolutional neural network which combines the advantages of recurrent neural networks and convolutional neural networks.It models temporal dynamics and spatial correlations simultaneously with the proposed cascade convolutional blocks where spatial features for different time steps are fused while their temporal orders are kept.This network first extracts spatial features layer by layer with ordinary convolutions,then uses cascade convolutional blocks to capture temporal dynamics in spatial features,and finally maps the predicted spatial features to original spatial correlation space with ordinary convolutions.The main contributions of this paper are as follows:· This paper is the first to propose a fully convolutional neural network for spatiotemporal predictive learning.The proposed network makes it available for this problem to get rid of recurrent models.· The proposed cascade convolutional block applies ordinary convolutions to abstract spatial features and utilizes its internal cascade structure to capture temporal features so that temporal dynamics and spatial correlations can be modeled simultaneously.· A direct link of future and historical input sequences is established built upon ascade convolutional blocks in the proposed deep cascade convolutional neural network and it avoids the similar chain structure in recurrent models.Therefore,the gradient vanishing problem is eased and a parallel optimization is enabled in the training process.The proposed deep cascade convolutional neural network not only achieves a significant relative improvement of 12% on the standard digital motion prediction task,but also is applied to two realistic scenarios of traffic flow forecast and air quality prediction.Experimental results show that deep cascade convolutional neural network is far superior to existing methods in various metrics such as root mean square error,model training time and memory usage,which proves its effectiveness and availability in realistic spatiotemporal predictive learning tasks.