Heat-Wave Prediction via Joint Regression and Classification with a Bayesian Deep Bidirectional LSTM Model

Abstract

Extreme heat-waves are severe climate hazards, threatening public health, ecosystems, and urban infrastructure. Accurate and uncertainty-aware prediction is vital for early warning, risk management, and climate adaptation. This study introduces a Bayesian Deep Bidirectional LSTM (BiLSTM) framework that unifies regression and classification for heat-wave prediction, enabling both maximum temperature forecasting and reliable event detection. The framework integrates Bayesian inference, bidirectional memory, and extreme value analysis for event labeling, while class imbalance is addressed using a Dirichlet Extended SMOTE method to generate distribution-aware synthetic samples, enhancing robustness under abnormal heat-wave patterns. Hyperparameters are optimized via Bayesian optimization, improving predictive stability. Results show strong performance with balanced accuracy of 0.93, F1-score of 0.8868, RMSE of 0.1829, R² = 0.89, and reliable uncertainty quantification (95% PICP = 93.31%). The proposed framework offers a scalable and reliable result for operational temperature forecasting with heat-wave detection, implication for climate planning and decision making. For practical deployment, a “Heat Wave Prediction App” was developed based on the proposed model provides seven-day forecasts and heat-wave detection.