Comparison of Machine Learning Algorithms for Performance Evaluation of Photovoltaic Energy Forecasting and Management in the TinyML Framework

Abstract

The availability of reliable photovoltaic (PV) power forecasting tools is an important factor for the dissemination of this technology. This is true not only for the integration of these difficult to predict sources in large power grids but also for small grids or standalone applications. The concept of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">edge computing</i> , through the use of small, low power and inexpensive devices can help to make predictions more localized and feasible also in small size applications. In this article prediction methods based on Artificial Neural Networks (ANNs) models are considered and compared, along with the possibility of reducing their cost in terms of memory and computational power requirements possibly without increasing prediction error. It is shown that quantization and pruning methods, implemented in the AI libraries of a common platform for Microcontroller programming, is a viable solution of this problem. Solar panel aging effects are also considered, and it is shown how the same system used for prediction can be an indicator of reduced plant efficiency.