An edited version of this article first appeared in pv magazine USA
As another year of extreme weather is hitting the world with a vengeance, the urgency of stemming the tide is profound. While there certainly isn’t a single solution to the problem, maximizing green energy production is a one of the things that can, hopefully, help us save the planet. Renewable energy is complex and, at times, unreliable, but Artificial Intelligence (AI) can help remove obstacles and unleash its true power.
In a previous article, we looked at how AI can help optimize wind power generation. But there is another green energy out there, often represented in yellow, that is infinite and widely available. If the sun is the driving force behind natural life and growth, if the sun is responsible for photosynthesis, it certainly has what it takes to generate electricity. The technology that proved this assertion to be right is photovoltaics (PV).
In the solar energy sector, just like in the wind sector, one of the areas where AI makes a serious difference, is forecasting, whether for production or demand. To predict output, AI uses predictive machine learning algorithms that are both model-based and data-driven systems. Solcast, an Australian solar forecasting vendor, offers a model that is comprised of an Artificial Neural Network (ANN), a complex, trained and multi-layered computing system whose architecture is directly inspired by that of the biological brain, to predict the hourly solar power generation of a given PV module.
In 2018, the Australian Energy Market Operator (AEMO), which had been solely responsible for predicting the country’s electricity production until then, announced that solar farms could “self-forecast” output. A number of large-scale solar sites in Australia have then partnered with Solcast to harness AI-based predictive technologies. Solcast employs a third-generation satellite “nowcasting” system that can detect and predict weather and sky characteristics. It geo-codes, quality controls and computes solar irradiance estimates by combining one kilometer resolution imagery from five geostationary weather satellites, complemented with later introduced real-time solar farm power data feeds and ground-based sky images of local cloud coverage. This contributes to 18 model predictions during each update and more than 600 million forecasts calculated per hour. The estimates are refreshed every 5 to 15 minutes and made available via the Application Programming Interface (API) in a matter of tens of seconds.
A year prior to AEMO’s announcement, Solcast’s AI forecasting test project had revealed compelling and solid results, yielding a 10 to 15% improvement over AEMO’s models. This application of AI in solar energy power production helps to better understand the sites and to improve productivity by reducing the need for frequency control services and better managing supply and demand across the National Electricity Market (NME). Ultimately, the goal of machine learning forecasting is to maximize accuracy in order to avoid unexpected fluctuations that can negatively impact grid stability and therefore consumer-end quality of life. Hence, significantly enhance system reliability and reduce electricity price.
Another field where AI turns out to be very useful is asset maintenance, which includes inspection and cleaning. Traditionally, solar farm maintenance would require a team of workers who would manually inspect the panels. The process can easily become tedious and expensive and cannot be consistently accurate or performed daily for an entire deployment. Operators have turned to AI to speed up the process and to improve accuracy, thus reducing the costs of quality control of entire facilities. Unmanned Aerial Vehicle (UAV) – or drones – are the most effective and common tools used by AI in solar to collect data and deliver aerial imagery. The captured images are then run through a deep learning algorithm trained to identify and classify visible signs of defects from a large quantity of previously processed datasets of labeled images.
Cloudy weather is not the only natural cause that can hamper solar energy production. PV panel arrays also suffer from the accumulation of dust, bird droppings, pollen, and other forms of particles that naturally form and travel outdoors. While rainy and windy weather conditions may sometimes remove some of the aggregated dirt, they often contribute to the opposite effect, a process called cementation. Morning dew and rainwater can add a layer of moisture on the panels that fosters the stiffening of the dust already present on the panels. This falls under the natural phenomenon of soiling, which hinders the penetration of sunlight, and can sometimes even damage components, thus shortening asset lifespan. A recent study estimated that the global cost of soiling could rise to over $6 billion by 2023. Scientists also proved that in some parts of the US, soiling could lead to a 7-percent energy loss per year, and up to 50% energy output reductions per year in the Middle East and parts of Asia.
Usually, solar farm operators would turn to soiling stations to solve this problem. These stations consist of a comparative energy loss measurement method using two sensors, one cleaned regularly and the other one untouched and therefore soiled, giving insight on the level and frequency of cleaning likely to be needed. Soiling stations imply extra costs to a project and don’t always take into account the variance of soiling levels within a single large-scale solar plant, nor do they necessarily anticipate precipitation that can contribute to the cleaning of panels. This leads to possible inaccuracy and unnecessary cleaning costs. AI vendors have now come up with solutions that can put minds worried about soiling mitigation costs to rest. Indeed, AI-powered platforms make use of the vast data streams that flow from solar panels to compare the effect of soiling and energy loss against actual costs of cleaning the panels depending on method and frequency. AI soiling mitigation solutions tap into data already available from assets (SCADA, historical weather data and forecasts, project-specific) and use machine learning to optimize maintenance tasks like cleaning, thus increasing profitability.
These are just a few examples of how AI and machine learning can help solar farms make their mark in the energy industry. But mostly, they are very significant examples of how AI can contribute to the urgent and imperative mitigation of fossil energy production and climate change.