A Comparative Study of Time Series Forecasting of Solar Energy Based on Irradiance Classification

Abstract

Listen

Translate article

Sustainable energy systems rely on energy yield from renewable resources such as solar radiation and wind, which are typically not on-demand and need to be stored or immediately consumed. Solar irradiance is a highly stochastic phenomenon depending on fluctuating atmospheric conditions, in particular clouds and aerosols. The complexity of weather conditions in terms of many variable parameters and their inherent unpredictability limit the performance and accuracy of solar power forecasting models. As renewable power penetration in electricity grids increases due to the rapid increase in the installation of photovoltaics (PV) systems, the resulting challenges are amplified. A regional PV power prediction system is presented and evaluated by providing forecasts up to 72 h ahead with an hourly time resolution. The proposed approach is based on a local radiation forecast model developed by Blue Sky. In this paper, we propose a novel method of deriving forecast equations by using an irradiance classification approach to cluster the dataset. A separate equation is derived using the GEKKO optimization tool, and an algorithm is assigned for each cluster. Several other linear regressions, time series and machine learning (ML) models are applied and compared. A feature selection process is used to select the most important weather parameters for solar power generation. Finally, considering the prediction errors in each cluster, a weighted average and an average ensemble model are also developed. The focus of this paper is the comparison of the capability and performance of statistical and ML methods for producing a reliable hourly day-ahead forecast of PV power by applying different skill scores. The proposed models are evaluated, results are compared for different models and the probabilistic time series forecast is presented. Results show that the irradiance classification approach reduces the forecasting error by a considerable margin, and the proposed GEKKO optimized model outperforms other machine learning and ensemble models. These findings also emphasize the potential of ML-based methods, which perform better in low-power and high-cloud conditions, as well as the need to build an ensemble or hybrid model based on different ML algorithms to achieve improved projections.