To address the issues of traditional BP neural networks being prone to local optima and exhibiting limited prediction accuracy in short-term station entry passenger flow prediction for urban rail transit, this study proposes an intelligent prediction method based on Improved Self-Adaptive Differential Evolution optimized BP Neural Network (ISADE-BPNN). First, the research object is defined as hourly station entry passenger flow, with analysis of its spatiotemporal characteristics and primary influencing factors. Second, a passenger flow data collection platform is constructed using Service-Oriented Architecture (SOA), and missing data imputation is performed based on probabilistic statistical methods. Third, to overcome the limitations of the conventional differential evolution algorithm, including fixed parameters and susceptibility to premature convergence, three improvements are proposed: an adaptive parameter adjustment strategy based on evolutionary state, a hybrid mutation strategy incorporating the spatiotemporal characteristics of passenger flow, and a learning mechanism based on historical success information. Finally, a two-stage hybrid prediction framework integrating “ISADE global optimization + gradient descent local fine-tuning” is constructed. Using metro lines from a first-tier city in China as the experimental subject, a theoretical analysis of the adaptive parameter mechanism is provided, demonstrating its convergence properties through monotonic evolution factor behavior and bounded parameter ranges. The proposed method is compared with GA-BPNN, DE-BPNN, and representative deep learning baseline models including Long Short-Term Memory (LSTM) networks and Gated Recurrent Unit (GRU) networks. Experimental results demonstrate that The proposed ISADE-BPNN method achieves improved prediction accuracy over the selected baselines in the hourly entry passenger flow task, outperforming the selected baselines of DE-BPNN, GA-BPNN, LSTM, and GRU across both RMSE and MAE evaluation metrics. Ablation studies with comprehensive computational experiments isolate the contribution of each improvement module, including the adaptive parameter strategy, and the Wilcoxon rank-sum test confirms that the performance differences between the proposed method and all comparison methods are statistically significant (p < 0.05).
Qian et al. (Fri,) studied this question.