Machine Learning-Based Solar Irradiance Prediction Using Low-Cost Environmental Sensors and ESP32

Abstract

Solar irradiance estimation is a key component in renewable energy management, precision agriculture, and environmental monitoring systems. However, conventional measurement instruments such as pyranometers often involve high acquisition and maintenance costs, limiting their deployment in low-resource environments. This work presents the development and implementation of a low-cost solar irradiance prediction system based on a Multilayer Perceptron (MLP) neural network embedded in an ESP32 microcontroller. Historical meteorological data collected from the NASA POWER database over a 13-year period were used to train the model. Temperature, relative humidity, and ultraviolet (UV) radiation were selected as input variables, while solar irradiance (W/m2) was considered the target output. The dataset was preprocessed through data cleaning and normalization procedures, and the MLP model was trained using the Adam optimization algorithm and ReLU activation functions. The resulting model achieved a coefficient of determination (R2) of 0.971, a root mean square error (RMSE) of 41.61 W/m2, and a mean absolute error (MAE) of 22.66 W/m2. After training, the model was converted to TensorFlow Lite format and deployed on an ESP32 microcontroller integrated with DHT11 and ML8511 low-cost sensors for real-time operation. Experimental validation performed under real operating conditions demonstrated an average prediction error close to 3% when compared with NASA POWER reference values. The results confirm the feasibility of implementing machine learning models on resource-constrained embedded platforms, providing an affordable and scalable solution for real-time solar irradiance estimation in energy and environmental applications.