“City-Aware Transformer-BiLSTM Model for Accurate Heatwave Prediction Using Daily Weather Data in Indian Cities”

Heatwaves are among India’s most critical climate-related hazards, significantly affecting public health, agriculture, energy systems, and urban infrastructure. Accurate and early prediction of heatwave events is therefore essential for reducing mortality, supporting disaster preparedness, and improving community resilience. This study investigates the effectiveness of advanced machine learning and deep learning models for forecasting heatwave occurrences using a comprehensive multicity daily weather dataset spanning 25 years (2000–2024). Data from ten major Indian metropolitan regions were used, incorporating key meteorological variables such as maximum and minimum temperature, apparent temperature, rainfall, humidity, wind speed, wind gusts, and wind direction. To capture temporal dependencies in weather patterns, a 28-day sliding window approach was applied. Data pre-processing included MinMax normalization, categorical feature embeddings, and SMOTE-based oversampling to address class imbalance. Two deep learning architectures were evaluated: a Transformer Encoder combined with BiLSTM, and a CNN-BiLSTM model. Their performance was compared with traditional machine learning methods, including XGBoost, LightGBM, SVM, Random Forest, and feed-forward neural networks. Results show that the Transformer-BiLSTM model achieved the highest accuracy (91.26%), outperforming the CNN-BiLSTM model (87.97%) and all classical approaches. The findings confirm that deep temporal sequence models provide superior performance for heatwave prediction. This study highlights the potential of deep learning-driven forecasting systems for strengthening early warning mechanisms and climate adaptation strategies in India.