Deep learning models for reliable weather forecasting in crowded regions performing traditional or religious activities

Forecasting weather during large-scale traditional or religious events in densely populated regions is particularly challenging because of the ever-evolving microclimatic changes caused by human activities. This study designed and tested two new deep learning-based approaches for automated weather predictions in such situations. A hybrid Convolutional Neural Network Long Short-Term Memory (CNNLSTM) framework is first proposed, which integrates forecasting based on the observation of spatial weather features using Convolutional Neural Networks for specialization in image analysis, and then employs Long Short-Term Memory networks to capture temporal dependencies. This model achieved 80% accuracy in precipitation prediction, with an F1-score of 0.80 for both the rain and no rain classes. The model is particularly strong in dealing with immediate microclimate impacts, such as crowd density and heat island features, which are extremely important for making accurate predictions. In the second part, we propose a spatiotemporal graph neural network (ST-GNN) that integrates weather conditions and crowd movement using a graph-based-attention mechanism. ST-GNN surpasses previously established forecasting models with an overall accuracy of 84.4%, F1-score of 0.843, and Receiver Operating Characteristics-Area Under Curve (ROC-AUC) value of 0.913. Detailed evaluations showed that high-risk situations were also significantly affected by extreme temperatures (+0.39 standardized difference) and high crowd density (+1.58 standardized units). Detailed explainability assessments of the models using SHapley Additive exPlanations (SHAP) and permutation importance identified shelter availability, crowd humidity effects, and the impact of warm active spaces within urban geography as the most important predictive features. Throughout the temporal changes, the models exhibited dynamic constancy in their performance, with precision-recall curves never falling below 0.85 in the training set. This study contributes to spatiotemporal deep learning in combination with practical solutions for event risk management by demonstrating that crowd data combined with meteorological information greatly enhance predictive accuracy in highly populated regions.