Regional Ground-Based IoT Solar Irradiance Monitoring: A Multi-Site Study Across Mountain, Rural, and Urban Environments

<jats:p>In this paper, the solar irradiance is investigated in a part of central Serbia, where a low-cost IoT sensor network was deployed at three locations: a mountain slope, an open field near a village, and an obstructed position in the city center. All three locations are in the same NASA POWER grid cell. After multi-stage quality control, 26,145 valid daytime records were compared to the satellite reference. The satellite assigns identical values to all three positions but the measured mean daytime irradiances are 267.7, 351.5 and 19.6 W/m2, respectively. The rural station is in the best agreement with the reference (R2 = 0.567); the mountain station suffers from a persistent shading bias (MBE = −138.9 W/m2); the signal at the urban station is attenuated by surrounding buildings and vegetation by a factor of ~12. Also, 25 regression models (gradient boosting, recurrent, convolutional, fully connected and graph-based) were trained on the 25-year monthly NASA POWER record for the same cell. XGBoost obtained R2 = 0.998, the hybrid TCN-GNN R2 = 0.968 and a plain ReLU network R2 = 0.912, and a seasonal model with calendar features was used to reconstruct the satellite reference for two months of 2026 not yet available in the archive. The deployed hardware, network and energy behavior are documented quantitatively: per-site link delivery ratios of 96.8%, 83.5% and 58.2%, the photovoltaic harvesting record of the nodes, and the absence of any energy-aware transmission scheduling. Two of the trained models were compiled for the ESP32 nodes and measured on the deployment hardware: a depth-limited gradient-boosted corrector runs in 46.3 µs using 11.1 kB of flash, and an INT8 fully connected network in 138.1 µs using 3.0 kB, while the graph hybrid cannot be converted for microcontroller execution at all. This defines an edge–cloud partition in which local inference performs bias correction and fault detection while the cloud tier retains the heavy models and periodic retraining. To the best of the authors’ knowledge, this is the first multi-site ground-based IoT irradiance record for central Serbia with such different terrain types.</jats:p>