A Comparative Analysis of Machine Learning and Traditional Time Series Forecasting Models to Predict Solar Energy Power

This research seeks to evaluate the comparative effectiveness of traditional time series models, such as ARIMA, against machine learning techniques in forecasting solar power output. Amidst observed deficiencies in the forecasting methodologies employed within Romania's solar energy sector, this study is imperative for understanding the potential of both conventional and novel predictive models to enhance sustainable energy development. Employing a univariate time series model (SARIMA) alongside a machine learning alternative support vector regression (SVR), this analysis leverages a dataset enriched with meteorological and temporal variables such as solar radiation, humidity, cloud cover, sunrise and sunset times, specific to the GPS coordinates of the investigated PV system. The production data for the 465.75 kWp PV system were collected using Huawei's FusionSolar monitoring platform. The results suggest that in terms of Romanian solar energy productions, machine learning models, specifically SVR, have outperformed traditional SARIMA models in forecast accuracy which supports the notion that there may be a shift to machine learning-based predictive models in the future. Using support vector regression models could greatly improve the accuracy of solar power forecasting in Romania, arguably resulting in better grid management and more efficient allocation of future resources. Furthermore, the results indicate that policy updates should be considered to support more advanced machine learning technologies in the sector of renewable energy.