Transformer based models with hierarchical graph representations for enhanced climate forecasting

Accurate climate predictions are essential for agriculture, urban planning, and disaster management. Traditional forecasting methods often struggle with regional accuracy, computational demands, and scalability. This study proposes a Transformer-based deep learning model for daily temperature forecasting, utilizing historical climate data from Delhi (2013-2017, consisting of 1,500 daily records). The model integrates three key components: Spatial-Temporal Fusion Module (STFM) to capture spatiotemporal dependencies, Hierarchical Graph Representation and Analysis (HGRA) to model structured climate relationships, and Dynamic Temporal Graph Attention Mechanism (DT-GAM) to enhance temporal feature extraction. To improve computational efficiency and feature selection, we introduce a hybrid optimization approach (HWOA-TTA) that combines the Whale Optimization Algorithm (WOA) and Tiki-Taka Algorithm (TTA). Experimental results demonstrate that the proposed model outperforms baseline models (RF-LSTM-XGBoost, cGAN, CNN + LSTM, and MC-LSTM) by achieving 7.8% higher accuracy, 6.3% improvement in recall, and 8.1% enhancement in F1-score. Additionally, training time is reduced by 22.4% compared to conventional deep learning models, demonstrating improved computational efficiency. These findings highlight the effectiveness of hierarchical graph-based deep learning models for scalable and accurate climate forecasting. Future work will focus on validating the model across diverse climatic regions and enhancing real-time deployment feasibility.