Accurate Short-Term Solar Irradiance Forecasting with TinyML on Edge Device

Accurate short-term forecasts of Global Horizontal Irradiance (GHI) are crucial for optimizing solar energy production and grid stability. They enable adjustments to power output based on anticipated fluctuations in solar power input. This paper proposes a novel Tiny Machine Learning (TinyML) approach for predicting GHI values specifically designed for deployment on resource-constrained Internet of Things (IoT) devices. Two experiments were conducted. First, a comprehensive benchmark study compared five machine learning models (artificial neural network, XGBoost decision tree, gradient boosting, and random forest) for GHI forecasting in Fez, Morocco. This study identified the ANN model as the most effective for simultaneously predicting GHI three multi-steps ahead. The benchmarking was performed on a PC to simulate cloud-based forecasting. The second experiment deployed and tested the chosen ANN model on a Raspberry Pi 4 to predict GHI three multi-steps ahead (1-3 hours). This evaluated its effectiveness for on-device (edge) prediction, contributing to a more intelligent and efficient solar energy management strategy.

I. Introduction

The increasing penetration of solar energy requires accurate and precise information about solar irradiance to maintain grid reliability [1]. However, solar irradiance data at ground level is inherently variable in time and space [2]. These weather and climate-driven fluctuations significantly impact solar power production [3]. Therefore, accurate knowledge of solar irradiance is crucial for optimal solar system design and operation [4]. Developing advanced technologies to manage the variability of solar irradiance and avoid economic losses is critical for making solar energy production more sustainable, reliable, and capable of achieving a continuous balance between energy demand and supply.