Abstract Vietnam ranks among the most climate-vulnerable countries in Southeast Asia, yet its operational weather monitoring network covers only approximately one station per 800 km², far below the World Meteorological Organization's recommended density. This paper presents the design, implementation, and field validation of a fully solar-powered IoT weather station network integrated with a Lambda/Kappa Big Data architecture and a Temporal Fusion Transformer (TFT) deep learning model for hyperlocal, multi-horizon weather forecasting. Fifty prototype stations were deployed across three climatically distinct provinces — An Giang (Mekong Delta flooding), Quang Nam (Central Vietnam typhoons/flash floods), and Dak Lak (Central Highlands drought/landslides) — collecting eight meteorological variables at six-minute intervals over twelve months. A Federated Anomaly Detection framework with differential privacy guarantees (ε = 2.0, δ = 10⁻⁵) identifies five classes of sensor faults without transmitting raw data. The TFT model, pre-trained on 23 years of ERA5 reanalysis and fine-tuned on local observations, achieves a 2-metre temperature RMSE of 0.81°C at 24-hour lead time, surpassing Kriging interpolation by 31% and the operational NWP baseline by 18%. System uptime reaches 99.3% over the evaluation period. All data and model weights are released as open datasets (CC BY 4.0) to support future research.
Bảo et al. (Sat,) studied this question.