Mastering the Urban Pulse: How Deep Learning Models Reshape the Prediction Accuracy and Resilience of Rail Transit Passenger Flow

Quantifying the Urban Pulse: From Complex Passenger Flow Prediction to Building Urban Transportation Resilience

China is undergoing an unprecedented urbanization process, and urban rail transit (URT) has become a core pillar in alleviating traffic congestion and achieving sustainable development. As of the end of 2023, the operating mileage of urban rail transit in China has exceeded 11,322 kilometers, with annual passenger volume increasing by over 51% year-on-year. This not only signifies the expansion of infrastructure but also highlights the structural challenges the urban transportation system faces under immense pressure.

However, this scale of expansion brings exponential growth in operational complexity. When passenger flow on certain key lines exceeds 130% of the design capacity, the system enters a critical state. This localized overload is not an isolated event; due to the complex topology of modern metro networks, a local congestion at one station can quickly trigger a chain reaction, threatening the operational stability and economic benefits of the entire network. Therefore, accurate short-term passenger flow prediction is no longer just an added feature but the cornerstone for urban transportation operators to dynamically allocate resources and implement proactive crowd diversion strategies.

Traditional prediction models based on Backpropagation Neural Networks (BPNNs), although applied in basic research, often reveal limitations when facing time-space dependent systems like rail transit, manifesting as an easy tendency to fall into local optima and difficulty capturing complex spatio-temporal dynamics. This directly constrains the operators' ability to make fine-tuned interventions during peak hours.

To bridge this gap, this study focuses the prediction target on "hourly station passenger inflow," deeply analyzing its spatio-temporal characteristics and main driving factors. The research path first constructed a data collection platform based on a Service-Oriented Architecture (SOA) and employed probabilistic statistical methods to handle data missingness, ensuring the quality of input data. Subsequently, the core innovation lies in the in-depth optimization of the Differential Evolution (DE) algorithm. The research team designed an innovative set of improvement strategies to address the inherent flaws of traditional DE algorithms, such as fixed parameters and insufficient convergence, including: an adaptive parameter adjustment mechanism based on the evolutionary state, a hybrid mutation strategy that incorporates passenger flow spatio-temporal characteristics, and a learning mechanism based on historical successful information.

Ultimately, the study constructed a two-stage hybrid prediction framework: "ISADE global optimization + gradient descent local fine-tuning." In experimental validation, compared against deep learning baselines such as GA-BPNN, DE-BPNN, and representative LSTM and GRU models, the proposed ISADE-BPNN achieved superior performance on both Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) metrics, with performance differences possessing significant statistical significance.From a macro perspective, this research is not just about optimizing an algorithm; it reveals that the combination of "model adaptability" and "algorithmic robustness" is key to enhancing system resilience in complex, dynamic urban infrastructure management. For urban researchers and policymakers, this means that in the planning and operation stages, this cutting-edge, high-precision predictive capability needs to be embedded into decision-support systems, thereby upgrading traditional passive emergency response into an active, intelligent passenger flow management system.

Looking ahead, as urban transportation networks evolve towards higher density and greater connectivity, the demands for accuracy and speed in predictive models will continue to increase. Future development directions may no longer be limited to the iteration of single algorithms, but rather towards the evolution of Self-Adaptive machine learning paradigms that can absorb new data streams in real-time and automatically optimize their search space. This foreshadows that urban intelligent transportation systems will shift from passive response to active evolution, truly achieving precise capture and effective regulation of the city's pulse.