Hyper-local weather forecasting has become increasingly critical due to rising climate variability and the precision requirements of modern agriculture and transportation systems. Conventional Numerical Weather Prediction models operate effectively at regional scales but struggle to capture microclimatic variations at fine spatial and temporal resolutions. This work presents an end-to-end hyper-local forecasting framework integrating IoT-based environmental sensing with a Transformer Encoder-based deep learning model. Multivariate time-series data comprising temperature, humidity, wind speed, solar radiation, and precipitation are processed using multi-head self-attention mechanisms to capture long-range temporal dependencies. Sinusoidal temporal embeddings are employed to model diurnal weather cycles, significantly improving short-term forecast accuracy. Experimental results demonstrate realistic transitions from high-radiation daytime conditions to nocturnal cooling and precipitation onset. The proposed system enables real-time decision support for precision irrigation planning and travel safety applications, validating its effectiveness for deployment in intelligent weather-aware systems.
Mahulkar et al. (2026) studied this question.