Improving solar forecasting using Deep Learning and Portfolio Theory integration

• Published in: Energy (Oxford) Vol. 195; p. 117016
• Main Authors: Lima, Marcello Anderson F.B., Carvalho, Paulo C.M., Fernández-Ramírez, Luis M., Braga, Arthur P.S.
• Format: Journal Article
• Language: English
• Published: Oxford Elsevier Ltd 15.03.2020; Elsevier BV
• Subjects: Alternative energy sources; Artificial intelligence; Deep Learning; Electric power generation; Electricity consumption; Energy demand; Energy sources; Forecasting; Integration; Machine learning; Multilayer perceptrons; Portfolio Theory; Predictions; Renewable energy sources; Solar energy; Solar forecast; Support vector machines; Topology; Brazil; Spain; National Institute of Meteorology; Mean Absolute Percentage Error; Photovoltaic Geographical Information System; Mean Absolute Error; Support Vector Machine; Coordination for higher Education Staff Development; Photovoltaic; Artificial Neural Networks; Recurrent Neural Network; Multilayer Perceptron; Radial Basis Function; Mean Percentage Error; National Council for Scientific and Technological Development; Statistical Evaluation; Forecast based on Portfolio Theory; Support Vector Regression; Aerosol Optical Depth; Institute for Energy and Transport; Root Mean Squared Error; Deep Learning; coefficient of determination; Solar energy; Long Short-Term Memory; Transmission System Operators; Artificial intelligence; Portfolio Theory; Solar forecast
• ISSN: 0360-5442; 1873-6785
• DOI: 10.1016/j.energy.2020.117016

Solar energy has been consolidated as one of the main renewable energy sources capable of contributing to supply global energy demand. However, the solar resource has intermittent feature in electricity production, making it difficult to manage the electrical system. Hence, we propose the application of Deep Learning (DL), one of the emerging themes in the field of Artificial Intelligence (AI), as a solar predictor. To attest its capacity, the technique is compared with other consolidated solar forecasting strategies such as Multilayer Perceptron, Radial Base Function and Support Vector Regression. Additionally, integration of AI methods in a new adaptive topology based on the Portfolio Theory (PT) is proposed hereby to improve solar forecasts. PT takes advantage of diversified forecast assets: when one of the assets shows prediction errors, these are offset by another asset. After testing with data from Spain and Brazil, results show that the Mean Absolute Percentage Error (MAPE) for predictions using DL is 6.89% and for the proposed integration (called PrevPT) is 5.36% concerning data from Spain. For the data from Brazil, MAPE for predictions using DL is 6.08% and 4.52% for PrevPT. In both cases, DL and PrevPT results are better than the other techniques being used. •Deep Learning exhibits better results than other consolidated solar forecasting techniques.•The integration of Artificial Intelligence methods in an adaptive topology based on the PT is a pioneering strategy.•The proposed integration reduces predictability errors.