Research On Super-resolution Inference And Prediction For Urban Flows

In recent years,with the progressive urban modernization process and the rapid development of intelligent terminals and wireless communication,a large amount of multimodal data springs into the urban Io T,providing intelligent services for people’s daily life while bringing challenges such as energy consumption increasing and public traffic congestion.To address these issues,existing work deepens the modeling of urban spatial structure and related external factors.Although these innovations bring considerable accuracy improvements,the drawback is that the information content of data itself limits the model fitting ability.To this end,we analyze the distribution pattern of coarse-grained historical observation and combine the internal structural features and external factors to obtain fine-grained urban data with richer feature profiles,which help alleviate the urban planning challenges such as irrational resource allocation and public traffic congestion.First,we propose an enhanced UrbanFlow-aware Super Resolution-Generative Adversarial Network(UrbanSG)for inferring fine-grained urban flow heat maps to guide the allocation of urban resources.Specifically,we use a conditional GAN as the backbone,which takes external factor as specific condition and finally generates finegrained inference which meet specific external conditions through an adversarial feature extraction process.To capture the implicit urban structure correlation,we introduce the Flow Self-Attention Module(FSAM)into the model,which focuses the model on traffic-active urban grid cells and effectively helps to handle sparse datasets.Secondly,we add a temporal dimension to the super-resolution perception framework and propose an end-to-end Spatio-Temporal Super Resolution UrbanFlow-aware Network(STRUrbaN)for urban traffic prediction,which predicts the fine-grained traffic under future time slices.STRUrbaN consists of two phases,data preparation and model prediction,in which the former selects key time slices among coarse-grained historical observations and fuses the corresponding external factors to learn a priori knowledge;the model prediction phase proposes a Temporal-aware UrbanFlow Encoder(TUFE)with Pre-Conv blocks and vision transformer to learn spatio-temporal dependencies in local and global contexts.In addition,we follow UrbanSG as the backbone to learn high-level feature representations of each urban grid cell and obtain the final fine-grained urban flow heat map.We conduct extensive experiments with two models on the public dataset and the Changchun mobile signaling dataset,respectively.First,the super-resolution accuracy experiment demonstrates that UrbanSG and STRUrbaN maintain a superior performance on both datasets compared with existing methods,especially in the sparse signaling dataset,bringing 13.87% and 11.64% performance improvement,respectively.Second,we Validate the effectiveness of each module(FSAM,TUFE,etc.)through multiple sets of ablation experiments,and hyperparameter comparison experiments to analyze and discuss the effects of different hyperparameter settings on the model performance.