Intelligent hyperlocal weather forecasting system based on small-sized neural network models and local data

Accurate short-term weather forecasting remains a key challenge in meteorology, particularly at the local scale where global numerical models (e.g., WRF, GFS) often fail to capture fine-grained atmospheric dynamics. This study presents a neural network–based approach designed to improve the accuracy of short-term (up to 24 hours) weather forecasts by integrating local meteorological observations with outputs from large-scale models. The proposed architecture follows an encoder–decoder structure with multi-head attention, enabling the model to learn spatiotemporal dependencies and correct systematic biases inherent in global predictions. To ensure real-time applicability, a lightweight TinyML version of the model was developed and deployed on an ESP32-S3 microcontroller, achieving inference latency below one second while maintaining comparable accuracy. Experimental results demonstrate a significant reduction in mean absolute error (MAE = 0.21 for temperature, 0.40 for pressure, 0.11 for wind speed) and high correlation with reference data (R² > 0.97). The proposed solution highlights the potential of integrating dense local sensing networks with resource-efficient neural models for hyperlocal weather forecasting, offering new opportunities for decentralized, on-device environmental intelligence in energy, construction, and infrastructure management.