Enhancing Photovoltaic Power Forecasting Using the LGB Model and Synthetic Features

Abstract

Over the past decade, the growing demand for sustainable energy has led to significant interest in photovoltaic (PV) power generation. Due to its intermittent nature, accurate PV power forecasting is essential for the efficient management and monitoring of PV systems. In this context, the accuracy of meteorological data is critical. However, it is not always possible to obtain such data on a local basis, and often some information, such as irradiance, is obtained from models whose characteristics are not known in detail. To overcome this limitation, this study evaluates five synthetic features that combine clear-sky global horizontal irradiance and cloudiness data to estimate total irradiance in the absence of direct measurements. A light gradient boosting model is used to evaluate the predictive performance of a model using these synthetic features compared to a model based on conventional meteorological inputs, including irradiance. The results, evaluated over a reference week, show that the feature labeled $\chi _{5}$ slightly improves model accuracy (passing from an RMSE of 84.013 to 87.232 W/kWp and $R^{2}$ from 0.888 to 0.875). These results show that synthetic features can achieve comparable results and in some cases even improve prediction performance.