Hybrid depth learning model LSTM-ResNet based metropolitan space-time flow prediction technology

Abstract

The invention relates to a hybrid depth learning model LSTM-ResNet based metropolitan space-time flow prediction technology. The invention can accurately predicting the change of urban spatio-temporaldata stream so as to provide important reference for urban management, and the key is to extract spatio-temporal dependency features from the data effectively. Currently, convolution neural network,which has been applied to spatio-temporal flow prediction, focuses on the extraction of spatial correlation features, ignoring the temporal dimension dependency and spatio-temporal correlation features. In depth learning model, long and short memory network (LSTM) is suitable for dynamic modeling of time series, and residual convolution network (ResNet) is suitable for large-scale spatial correlation feature extraction. Therefore, we combine LSTM and ResNet to construct a hybrid depth-learning model for spatio-temporal flow prediction: LSTM is used to consider the time dependency before and after, and filter out the invalid time features; the output of LSTM is inputted into ResNet and the spatio-temporal correlation feature is extracted. The model can automatically and accurately capture spatio-temporal correlation features, especially retaining valid temporal features when considering forward and backward dependencies.

Description

technical field [0001] The invention belongs to the technical field of geographic information, and in particular relates to a prediction technology of spatio-temporal data flow on a city scale. Background technique [0002] In our daily life, people interact spatio-temporally with urban spaces through various behavioral activities (such as driving, cycling, walking, etc.). In recent years, thanks to rapid advances in sensor technology and the internet, a wealth of mobile data generated by these activities can be recorded. Usually, a typical mobility dataset consists of a group of objects (such as people, private cars, or buses) and their trajectories in space and time, which contains rich spatiotemporal information. By summarizing these mobile data sets according to time and space dimensions, spatio-temporal flow data can be obtained, which usually includes two basic types: input flow and output flow. For a given spatial unit, the number of objects entering this spatial un...

AI Technical Summary

Problems solved by technology

Although the spatio-temporal residual network (ST-ResNet) considers temporal and spatial features simultaneously to a certain extent,

Images