Predicting Extreme Weather: A Theoretical Semi-Supervised Machine Learning Framework | RISE Research

This paper presents a meta-review of extreme weather prediction methods, spanning physics-based numerical models, classical machine learning, and modern AI systems, drawing on five key studies from 1992 to 2025, before proposing two original theoretical models to address identified gaps. The literature review finds that global climate models operate at too coarse a spatial resolution for local decision-making (such as predicting a flash flood in a specific drainage basin), that model accuracy declines for fast-changing dynamic events like high winds compared to slower thermodynamic events like temperature, that computational costs remain prohibitive for high-resolution multi-dimensional simulations, and that many AI-driven models focus on pattern recognition without capturing underlying physical causes, risking bias, especially given limited region-specific data for rare extreme events. In response, the paper proposes Model 1, a spatiotemporal gradient boosting framework using Gaussian weighted downscaling, which assigns higher weight to training data from cities and time points closer to the prediction target (a 'smooth fade' approach) rather than treating all data uniformly, aiming to improve resolution-aware forecasting; and Model 2, a semi-supervised feedforward neural network that first pretrains on a large pool of unlabeled meteorological observations to learn normal atmospheric patterns, then fine-tunes on a smaller labeled dataset to distinguish extreme from normal weather conditions, aiming to improve rare-event detection while remaining computationally lightweight enough for real-time use. Both models are theoretically contextualized against a publicly available multi-city hourly weather dataset covering 36 countries, though the paper presents architectural design and rationale rather than full empirical validation. The paper concludes that continued research and testing of these frameworks could improve extreme weather prediction accuracy and deployability, ultimately supporting public safety and disaster preparedness.