Author(s): Silva, Rodrigo Amaro e
Date: 2020
Persistent ID: http://hdl.handle.net/10451/47449
Origin: Repositório da Universidade de Lisboa
Project/scholarship: info:eu-repo/grantAgreement/FCT/OE/PD%2FBD%2F106007%2F2014/PT;
Subject(s): previsão solar; previsão solar espácio-temporal; solar forecasting; spatio-temporal solar forecasting; Domínio/Área Científica::Ciências Naturais::Ciências da Terra e do Ambiente
Current and future photovoltaic (PV) deployment levels require accurate forecasting to ensure grid stability. Spatio-temporal solar forecasting is a recent solar forecasting approach that explores spatially distributed solar data sets, either irradiance or photovoltaic power output, modeling cloud advection patterns to improve forecasting accuracy. This thesis contributes to further understanding of the potential and limitations of this approach, for different spatial and temporal scales, using different data sources; and its sensitivity to prevailing local weather patterns. Three irradiance data sets with different spatial coverages (from meters to hundreds of kilometers) and time resolutions (from seconds to days) were investigated using linear autoregressive models with external inputs (ARX). Adding neighboring data led to accuracy gains up to 20-40 % for all datasets. Spatial patterns matching the local prevailing winds could be identified in the model coefficients and the achieved forecast skill whenever the forecast horizon was of the order of scale of the distance between sensors divided by cloud speed. For one of the sets, it was shown that the ARX model underperformed for non-prevailing winds. Thus, a regime-based approach driven by wind information is proposed, where specialized models are trained for different ranges of wind speed and wind direction. Although forecast skill improves by up to 55.2 % for individual regimes, the overall improvement is only of 4.3 %, as those winds have a low representation in the data. By converting the highest resolution irradiance data set to PV power, it was also shown that forecast accuracy is sensitive to module tilt and orientation. Results are shown to be correlated with the difference in tilt and orientation between systems, indicating that clear-sky normalization is not totally effective in removing the geometry dependence of solar irradiance. Thus, non-linear approaches, such as machine learning algorithms, should be tested for modelling the non-linearity introduced by the mounting diversity from neighboring systems in spatio-temporal forecasting.
