Evaluation of machine-learning-based solar PV and wind power regional models for Spain

The share of solar/wind energy in the electricity systems of many countries in the world will reach unprecedented values in the coming decades, fostered by the mitigation of climate change and also by the economic competitiveness of these energies. Accurate regional models of wind/solar power generation are crucial to making a smooth transition to these new energy systems.

In this work, regional solar PV and wind power models for Spain, based on the Random Forest machine learning model, were built and evaluated. Models were obtained for each of the 50 Spanish regions using as input the installed wind/solar capacity of the corresponding region and a set of meteorological variables. Regional installed capacities values were elaborated based on information from wind farms at 506 locations and photovoltaic solar plants at 3854 locations collected from public databases. The study is carried out for the year 2018 at hourly resolution; the model estimates were evaluated based on the actual power generation values provided by the Spanish TSO. This study is part of the Spanish MET4LOWCAR project, that aims at demonstrating the benefits of design low carbon power systems that accounts for the regional climatic patterns of both solar and wind renewable resources, using the Spanish territory as a testbed.

Different studies were undertaken. Firstly, two different modeling approaches were evaluated. In the first one, the meteorological inputs were derived by simple averaging the values at all the grid cells of the corresponding region. In the second approach, meteorological inputs were computed as a weighted average at the solar/wind farms locations using as weights the corresponding installed capacity. Secondly, two different meteorological databases were evaluated: the ERA-5 reanalysis and specific database derived from an ad-hoc integration conducted with the WRF NWP model. This late database has a spatial resolution of 5 km and temporal resolution of 10 minutes. Derived from these databases, up to 11 meteorological variables were used as inputs for the models The relevance of these variables in the models performance was assessed based on the feature importance provided by the Random Forest modeling procedure.

Results show, firstly, that the two modeling approaches perform similarly, being the overall power estimates relative errors of about 20% for Solar PV and 30% for the Wind power. Secondly, that the models performance using the two meteorological databases are also similar, although the WRF databases provide slightly more accurate estimates of the power variability. Thirdly, the most important meteorological input variables for the solar models were the GHI, DNI and temperature, while for the wind models were the wind at 10 and 100 meters above the ground and the wind at the 850 hPa level. The sources of the uncertainty of the power estimates are discussed.