Neural network quantification for solar radiation prediction: An approach for low power devices

Abstract

Accurate solar radiation prediction leverages various machine learning techniques, with artificial neural networks (ANN) being the most common and precise due to their ability to detect and learn relationships between meteorological variables and solar radiation. Traditionally, training and deploying these models require high-capacity computers. However, the proliferation of low-power smart devices, such as embedded systems and mobile devices, necessitates exploring methodologies for implementing ANN on systems with limited computational resources. This paper proposes a quantized neural network model for solar radiation prediction, considering the hardware limitations of low-power devices like the Raspberry Pi RP2040 microcontroller. The methodology involves five stages: hardware and software selection, neural network development and quantization, microcontroller implementation, model validation, and result analysis. Experimental design allows detailed performance evaluation of quantized neural networks, demonstrating that the TensorFlow Lite Quantized Aware model is suitable for solar radiation prediction. Metrics such as root mean square error (RMSE) of 44.24 and R² of 0.96 indicate that the selected quantized model differs from the original non-quantized model by less than 0.5% in RMSE and 0.04% in R². The study concludes that implementing quantized ANN models on microcontrollers is a technically and economically viable solution for solar radiation prediction. Quantization enables complex predictive models to run on low-cost, energy-efficient devices, thereby democratizing advanced prediction technologies for critical applications like solar energy generation.