A covariate aware and dual convolutional network model for stadium crowd flow prediction

Campus physical activities and sports facility utilization exhibit significant temporal fluctuations, which directly impact the efficiency of teaching resource allocation and the effectiveness of student physical exercise. During peak hours, overcrowded sports venues not only compromise teaching quality but may also increase the risk of sports injuries, thereby undermining the sustainability and enthusiasm of student participation. Accurately predicting student participation trends across different time slots and venues is thus of great practical significance for optimizing physical education schedules, enabling staggered usage, and enhancing overall classroom efficiency and training outcomes. To address this challenge, we propose a crowd flow prediction model tailored for physical education scenarios. The model integrates a covariate-aware cross-attention mechanism and a dual-layer convolutional feedforward network (DConvFFN) to improve prediction accuracy and robustness. By incorporating external covariates such as time, temperature, and academic schedules, the model dynamically captures implicit relationships between participant flow and multiple influencing factors. The cross-attention mechanism extracts interactive information between temporal features and external variables, while the dual-layer convolutional structure enhances the model's capacity to learn complex feature dependencies and localized patterns. Experimental results demonstrate that the proposed model effectively identifies complex periodic patterns and sudden fluctuations in crowd trends within physical education settings. Compared to traditional Recurrent Neural Networks (RNN) and Transformer models, it reduces the Mean Squared Error (MSE) and Mean Absolute Error (MAE) by 6.16% and 4.06%, respectively. These results highlight the model's potential for practical applications in pre-allocating teaching resources and dynamically adjusting curricula, providing key technical support for building a data-driven smart sports teaching management system.