Development and Evaluation of Solar Radiation Sensor Using Cost-Effective Light Sensors and Machine Learning Techniques

Journal: Technologies (MDPI), published May 2025 (exact publication day not independently confirmed).

Key findings

An IoT-based system integrating low-cost ambient light sensors with cloud storage/processing, benchmarked against a dedicated Davis 6450 pyranometer reference and validated against meteorological API data. A random-forest regression model correcting raw BH1750/VEML7700 lux readings achieved R² = 0.9922 against the reference instrument — far better than simple linear regression, which one companion figure put at ~164 W/m² RMSE (roughly 54% error).

A commonly cited simple linear conversion factor is roughly lux × 0.079 ≈ W/m², but this is only approximate and varies with sun angle, cloud cover, and spectral composition; accuracy is better (~±4%) near solar noon on clear days.

Why it matters for this archive

Confirms that a $1–2 BH1750 lux sensor is a defensible “good enough” solar-irradiance proxy for a hobbyist environmental-monitoring build — with the important caveat that raw lux-to-W/m² conversion is rough without a correction model, and closer to true irradiance measurement requires either a machine-learning correction or a purpose-built photodiode sensor. See Small-scale passive-thermal experiment monitoring.