Research on temporal and spatial feature extraction and distribution prediction of the population in Lujiazui Financial & Trade Zone based on heat map

Considering the densely populated area of Lujiazui Financial & Trade Zone, we propose a population temporal and spatial distribution model based on heat map The collected spatiotemporal heat data are analyzed according to the distribution characteristics of time domain (statistical domain) and space domain (statistical domain) we propose four eigenvalues: total number, maximum, standard deviation and average for stationary and periodic analysis, and extract the feature vector in time domain through MLP. In spatial domain, we propose three different sizes of convolution to convolute spatiotemporal data to extract spatial feature vectors. Through the attention mechanism, the characteristics of time domain and space domain are fused by the weight of autonomous learning, and the spatio-temporal characteristics are obtained. A three-layer convolution neural network based on spatiotemporal fusion characteristics is constructed to predict the spatiotemporal distribution of population heat. After experiments, the model proposed in this paper is lower than other baseline models in error level. According to the attention weight of spatio-temporal features, the following conclusions are drawn: (1) in the time-domain features, the maximum and average can best represent the time-domain features. (2) in the spatial domain features, the larger the scale of convolution kernel, the better the prediction effect. (3) the periodicity of time-domain can be used as the time window of convolution.