Weather Forecasting System “Four Seasons in One Day” and Shelter Suggestion for Sydney and Melbourne and Canberra
摘要
This study aims to address rapid intra-day weather variability in Australian cities by developing a multimodal forecasting system. The proposed framework integrates meteorological data from the Bureau of Meteorology, satellite imagery (MODIS/Sentinel-2), and social media signals from Reddit, processed using a Swin Transformer and RoBERTa models. The Transformer-based WeatherNet achieves a Mean Absolute Error (MAE) of 0.13 \(^\circ \) C for temperature predictions and an AUC of 0.98 for rainfall probability. The system provides shelter recommendations and real-time alerts, enhancing urban safety and decision-making. Four predictive models are evaluated: ARIMA, XGBoost, Bi-LSTM, and a Transformer-based WeatherNet. WeatherNet achieves the best results, with a Mean Absolute Error (MAE) of 0.13 \(^\circ \) C and an AUC of 1.00. To ensure transparency, the system incorporates SHAP, Integrated Gradients, and DiCE explanations. It also recommends nearby shelters using OpenStreetMap points of interest. A fully automated pipeline, powered by agent-based modules and the n8n workflow engine, delivers daily forecasts and real-time alerts via dashboards. Ultimately, the system aims to enhance public safety and health by providing timely warnings and accessible shelter suggestions during extreme weather conditions.