Spatial–Temporal based Crowd Flow Prediction Using Multiscale Dynamic Residual Learning

ABSTRACT Accurate and immediate crowd flow prediction remains difficult mainly because it depends on recent crowd flow patterns and nearby locations. Current research focuses on space–time dependency modeling yet neglects how crowd movements link distant areas. Every period of weekly intervals shows repeating patterns, which define the periodic nature of crowd flow data. Current research about capturing cyclic periodicity faces challenges and incorporates it into network channels to boost strategies within the network structure. This paper introduces the Multiscale Residual Network (MSRNet) model to address the inadequacies in periodicity modeling for crowd‐flow data. The MSRNet framework presents crowd flow prediction through parallel periodic learning, where it compares past sequence segments with future segments for simultaneous modeling of cyclic changes. The natural dynamic nature of crowd flows makes direct prediction of their flows inherently tricky. The challenge of dynamic complexity requires trainers to concentrate on stationary fluctuations because these represent the most suitable way to train the model. We build a Multiscale Channel Encoder (MSCE) system using two connected components to find shared space–time features when modeling crowd inflow and outflow prediction as related tasks. The network establishes parallel connections with highways through nested long‐short‐term memory (NLSTM). The connection between future data points and their weekly observations improves predictive accuracy multistep ahead. The model performance evaluation depends on experimental diligence that contrasts its results with established models using two real‐world data collections.