BiLSTM-RiskNet: Domain-Guided Explainable Deep Learning for Multi-Class Climate Risk Prediction from Hourly Weather Data

Precise and transparent forecasting of climate risk levels remains a formidable task in disaster preparedness because most current approaches either utilize static rule-based thresholds or do not tackle temporal trends in weather observations. These approaches typically prioritize the accuracy of predictions at the expense of interpretability and are thus unsuitable in high-stakes early-warning systems. To fill such gaps, this paper presents BiLSTM-RiskNet, a new domain-knowledge-guided and explainable deep learning approach for climate risk classification (low, medium, high) from raw hourly weather time series. Compared to existing research, the model employs physics-constrained automated labeling according to authoritative meteorological guidelines like NOAA Heat Index, Saffir–Simpson wind scale, and WMO barometric pressure criteria to provide semantically meaningful supervision in place of random thresholds. A Bidirectional LSTM is used on 24-hour rolling sequences to identify forward and backward temporal trends, so the model is better able to identify changing risk patterns than fixed or unidirectional methods. To address deep learning's black-box issue, explainable AI techniques are incorporated: LIME supplies local, instance-level interpretation, and SHAP supplies global feature attribution. These descriptions corroborate the model's reasoning to be theoretically rooted in climatological theory, e.g., extreme heat, humidity, and low pressure forming a part of the factors that contribute towards risky conditions. Equipped with a stratified, city-conscious dataset of more than three million hourly observations, BiLSTM-RiskNet achieves 95.23% accuracy with F1-scores of 0.98 (low), 0.72 (medium), and 0.92 (high) surpassing traditional machine learning and deep learning baselines. This synthesis of physics-based labeling, temporal deep learning, and open post-hoc explainability is a solid and reliable paradigm for real-time operational deployment, and represents an important milestone in AI-based climate risk reduction and disaster resilience.