Short term solar irradiance forecasting via a novel evolutionary multi-model framework and performance assessment for sites with no solar irradiance data

• Published in: Renewable energy 2020-09, Vol.157, p.214-231
• Main Authors: Marzouq, Manal, El Fadili, Hakim, Zenkouar, Khalid, Lakhliai, Zakia, Amouzg, Mohammed
• Format: Article
• Language: English
• Subjects: Climatic zoning; Evolutionary ANN; Global solar irradiance; Machine learning models; Multi-model framework; Short term forecasting
• ISSN: 0960-1481; 1879-0682
• DOI: 10.1016/j.renene.2020.04.133

Accurate forecasting of solar irradiance is a key issue for planning and management of renewable solar energy production technologies. The present paper aims to propose new machine learning forecasting models based on optimized ANNs in order to accurately predict solar irradiance. For this purpose, an evolutionary framework is suggested to generate multiple models for different time horizons up to 6 h ahead by the evolution of the forecasting history and ANN architecture. A dataset of 28 Moroccan cities is used in our experiments in order to explore the performances of the proposed models against different climatic conditions. The proposed framework is then evaluated through a zoning scenario giving the ability to our models to accurately forecast solar irradiance in sites where no such data is available. Two other scenarios are used to assess and compare the resulting performances. For all studied scenarios obtained results show good generalization abilities with NRMSE varying from 7.59% to 12.49% and NMAE from 4.41% to 8.12% as best performances for solar irradiance forecasting from 1 to 6 h ahead respectively. A comparative study is then conducted with three other models (smart persistence, regression trees and random forest), showing better performances of our proposed HAEANN models. •New models based on optimized ANNs to forecast solar irradiance are proposed.•Evolutionary framework is proposed to forecast solar irradiance up to 6 h ahead.•A useful zoning scenario where no solar irradiance data are available is investigated.•Comparative study is performed with a naive model and machine learning models.•Proposed models outperformed the other evaluated machine learning models.