Up to now nobody really knew how many solar systems there are on earth, least of all where they are actually located. To find out, researchers from England and the USA combined for the first time artificial intelligence – machine learning and big data analysis – with satellite remote sensing.
During the inventory phase between 2016 and 2018, they discovered 68,661 photovoltaic structures in 131 countries with capacities of more than 10 kilowatts worldwide. That was 432 percent more than the best and most reliable estimates of international organizations would suggest. That would correspond to a total capacity of 350 to 500 gigawatts, calculated the scientists working with Lucas Kruitwagen from the University of Oxford.
With their estimate of the worldwide installed solar power of around 420 gigawatts for the end of 2018 were the International Energy Agency (IEA) and the International Renewable Energy Organization (IRENA) so not so wrong at all. That is almost twice as much as all power generation systems in Germany do today. At the end of 2018, 45.5 gigawatts of solar power were installed in this country, mainly as small and micro systems on roofs of private homes, which were not even counted in the global inventory campaign. In contrast, appreciated the World Resource Institute in Washington only 3.9 gigawatts of solar power for Germany at the time, IEA and IRENA came to 38.7 gigawatts.
Closing the knowledge gap on photovoltaic systems
Exact statistics on the spread of regenerative energies are rarely available, at best rough estimates that are far apart. In 2018, the World Resource Institute assumed 54.3 gigawatts installed in 5,289 plants worldwide, the world database of the electricity companies came to 12,915 plants with a total of 107.4 gigawatts. The IEA and IRENA estimates were more realistic, but how many plants existed and where they were located was unknown. Volker Quaschning (Professor for Renewable Energy Systems, University of Applied Sciences HTW Berlin) has the Development of photovoltaics based on estimates by IEA and IRENA summarized in tabular form over the last 30 years.
The scientists from Oxford, in collaboration with colleagues from the spatial data company Descartes Labs in Santa Fe, New Mexico, and the World Resource Institute, have now been able to close this huge knowledge gap with their combination of earth observation and machine learning. The results show that solar energy is booming primarily in Europe, but also in India, China, Japan and on the coasts of North America. In the sunny countries of Africa and Latin America, but also in Australia, solar systems are hardly widespread.
For politicians, climate negotiators, authorities, infrastructure and network planners all over the world, such a high-resolution database is a good tool for tracking and, above all, controlling the acceptance and spread of solar energy. The choice of land for the construction of renewable plants should play a much bigger role because it has an impact on land cover and changes in land use, which in turn affect greenhouse gas emissions. The wrong choice of location can also restrict food production or drive out indigenous peoples. So while solar systems contribute to the UN goals for sustainable development in the areas of clean energy, economic growth, infrastructure and climate protection, they can on the other hand harm goals such as fighting hunger, health, reducing inequality and living in the countryside.
Photovoltaic systems often operate on sensitive ecosystems
The satellite images also showed the researchers patterns of land cover that they found with older images from the European Space Agency (ESA) climate initiative compared. So they could see that most of the photovoltaic systems were built on farmland. In second place came dry areas and grasslands, both of which are extremely sensitive ecosystems. This also applies to Europe, with Germany and France taking more built-up areas into account.
However, the technical effort required to follow the expansion of photovoltaics in this way around the world is enormous. Lynn H. Kaack from the Hertie School’s Data Science Lab in Berlin made a comment because also very cautiously: “The costs and benefits of such approaches for society should be regularly assessed and compared with the alternative option of direct data collection.” However, it does acknowledge that initial fears that machine learning would require too much effort and provide insufficient accuracy have largely been invalidated by this initial success.
Deep learning models and manual labor
Obtaining high-resolution satellite images is extremely expensive. The researchers used the images from two earth observation satellites, the paired European Sentinel-2 system and the French SPOT. The quality of the images was nevertheless very different, so that thousands of images had to be labeled manually to train the artificial intelligence. Only then were the scientists able to develop deep learning models that could be used to analyze location, size, installation date and soil conditions. Manual manual work and eye work was again the order of the day after the test phase, for example to sort out chicken coops or greenhouses that the algorithms had incorrectly counted at the beginning.
In total, the mainframe of the Descartes Lab in Santa Fe had to process 550 terabytes of image material, including the central processing unit (CPU) 106 hours and the graphics processors 20,000 hours. That added up to an energy consumption of 71 megawatt hours, the annual consumption of 30 average single-person households in Germany.
The successful combination of earth observation and artificial intelligence makes it clear, however, that photovoltaics is a worldwide, often hidden, flourishing success story. But also that their growth must be planned far better so as not to thwart other transformation and development goals.