PERSONALISED WEATHER INTELLIGENCE SYSTEM: A HEALTH-AWARE MACHINE LEARNING APPROACH TO PERSONALISED WEATHER RECOMMENDATIONS

Abstract

This paper presents the design, implementation and evaluation of a Personalised Weather Intelligence System for health-aware outdoor decision support. The central problem addressed is that conventional weather applications present the same meteorological information to all users, even though temperature, humidity, air quality and ultraviolet exposure have different practical meanings for older adults, people with asthma, cardiovascular disease or diabetes, and users with high heat or cold sensitivity. The proposed system combines live data from the OpenWeatherMap API with individual health-profile inputs and trained machine learning models to generate four personalised outputs: adjusted perceived temperature, health-risk classification, activity-suitability assessment and forecast-based activity timing guidance. The machine learning pipeline used the ASHRAE Global Thermal Comfort Database II and retained 16,141 complete records from 109,033 original measurements after strict cleaning. Three predictive components were selected through systematic algorithm comparison: a Random Forest Regressor for perceived temperature prediction, a Random Forest Classifier for activity suitability, and a Neural Network Multi-Layer Perceptron for health risk classification. The final prototype was implemented using FastAPI, scikit-learn, joblib and a CDN-based React frontend. Results show strong performance for perceived temperature prediction (MAE = 0.216°C), activity suitability classification (weighted F1 = 0.968) and health risk classification (accuracy = 0.992). However, the health risk model’s high-risk recall of 0.719 confirms that the system remains a decision-support prototype rather than a deployable medical or safety-critical product. The paper concludes that personalised weather intelligence is technically feasible and practically meaningful, but future work requires outdoor thermal comfort data, class-balancing for high-risk examples, expert clinical validation and formal user testing. Its contribution is therefore both practical and critical: it demonstrates an integrated system while identifying the data, safety and user-evaluation requirements that must precede real deployment.