Low-Cost Hybrid Pyranometer with Machine Learning-Enhanced Solar Irradiance Sensing

Accurate measurements of solar radiation are essential for photovoltaic (PV) and other solar energy applications. However, standard thermopile pyranometers are expensive, which limits their deployment in a distributed network for these applications. In this study, a low-cost hybrid pyranometer that combines optical and thermal sensing with machine-learning techniques for high-accuracy solar irradiance estimation was developed. The system integrates a BH1750 digital light sensor and two thermistors, interfaced with an ESP32 microcontroller for data acquisition. To mitigate saturation under high irradiance conditions, the BH1750 was enclosed in a translucent Mylar-coated housing. One thermistor was attached to a black anodised aluminium absorber to sense solar-induced thermal effects, while a second one measured ambient-induced thermal effects. The system was deployed outdoors and synchronised with a Class-A XS-TBQ02A thermopile pyranometer as the reference pyranometer. The data were recorded at one-second intervals and comprised 57,000+ observations under varying meteorological conditions. It was then feature engineered to generate 40 candidate predictors from optical and thermal measurements, including temporal, statistical, derivative, interaction, and polynomial features. Least Absolute Shrinkage and Selection Operator (LASSO) analysis reduced the feature set to 26 informative predictors. Five machine-learning algorithms-Generalised Linear Regression (GLR), Support Vector Regression (SVR), Random Forest (RF), Multilayer Perceptron Neural Network (MLPNN), and Gradient Boosting (GB) were evaluated. Among these models, GB delivered the best performance, achieving an RMSE of 10.597 W/m², MAE of 5.162 W/m², R² of 0.9986, nRMSE of 2.3%, and MBE of 0.278 W/m², reducing prediction error by 80.23% relative to the best single-sensor baseline using an optimal ensemble size of 600 learners. The system can be fabricated for approximately US$14.60, a 95% cost reduction compared with standard pyranometers. These results demonstrate that inexpensive sensors combined with machine learning can provide highly accurate irradiance measurements suitable for solar energy applications.