Powering Solar Asset Management with Machine Learning
Around 2018, the overall cost of generating electricity from Renewable sources (solar, wind) became cheaper than the traditional methods of electricity generation (coal, oil, gas, nuclear). More than half of new electricity generation capacity added in 2021 were Renewables, and at the same time, the amount electricity distribution grids were willing to pay per unit of Renewable energy began to drop significantly. Managing the accelerated growth in capacity, while driving down costs, has become a must for Renewable plants.
Just as Renewable energy has grown in the last decade so has the field of Artificial Intelligence (AI). Traditional computing is software programmers creating algorithms, to solve for complex engineering problems. Whereas artificial intelligence (AI) is broadly computers capable of solving general problems, including those of reasoning, awareness, perception, sometimes further sophisticated concepts like empathy and compassion.
One subfield of AI, machine learning (ML), has witnessed particularly substantial progress over the years. In ML – the machine, (rather than a human programmer) is able to derive its own reusable algorithm after analysing massive quantities of data and outcomes. It’s remarkable how often these machine-derived algorithms are strikingly superior to the best-known programmer-designed algorithms.
Fortunately, ML applications applied to solar are helping operators sustain thriving business models which will only increase as algorithms improve, and solar plant operators better understand ML’s advantages from a time and financial point of view.
Solar plants have four characteristics that are challenging for human management and traditional computer algorithms.
Challenges, and ML solutions
The largest solar plants are now beyond 2GW in generating capacity. Even a mid-sized 200MW plant has about one million solar panels. While most of these panels work as expected, the human effort required to manually detect the small percentage that are underperforming, has become prohibitively expensive. ML anomaly detection algorithms, like autoencoders and GANs, are exceptional at identifying the few components behaving unusually among a large group. Using ML algorithms allows plant managers to accurately identify panels needing attention, then deploy their limited, expensive human resources to investigate potentially problematic panels.
At the national level, about a terawatt of electrical power is needed; this makes many geographically distributed plants unavoidable. Differing weather and local geography mean each plant behaves differently. ML learning algorithms have a transfer learning capability, which enables algorithms derived at one location to be used in alternate locations. This enables cheaper remote / centralized monitoring of plants, while each plant uses a locally adapted model.
Solar plants have about thirty distinct faults impacting energy generation. Certain faults (like shading) arise immediately upon installation, few like tracker failures and connection degradation occur after a few years, whereas faults like light-induced-damage and corrosion may take up to ten years to emerge. Computer programmer generated algorithms are static – only ideal for one stage of aging. ML algorithms, on the other hand, automatically modify their anomaly detection and performance prediction characteristics as plant behaviour changes. Additionally, as conditions vary (e.g., influence of climate change), these algorithms adapt to model the plant’s altered normal performance. This removes model maintenance and modification cost for plant operators.
Intermittency of operation
Solar energy is generated intermittently, with strong dependence on highly variable weather conditions for energy yield. At local plant level, this problem can be addressed by including storage (typically batteries). ML algorithms can analyse changing data about plant electricity generation and energy demand, then generate appropriate schedules for charging and discharging.
ML has already proved its worth in resolving problems associated with error detection and prediction, operational scheduling, and energy generation prediction to the extent that by 2025 over 80% of commercial operations are expected to be using ML. This is critical, as while the cost of constructing a solar plant is falling substantially each year, the cost of operations is now about half the overall cost.
Some considerable challenges remain
Integration of Long-Term Storage
This proportion of electricity generated by Renewable sources is expected to reach 56% in 2050. Seasonal variations in the amount of energy produced means that grid systems need to integrate long-term energy storage. Generating long term weather forecasts, and then predicting the plant’s corresponding energy production, ML can generate accurate predictions of the long-term energy generation. This provides capacity modelling for long-term storage (like pumped hydro), and scheduling usage of long-term storage to enhance energy availability.
Currently national electricity grids enable one-way flow of energy from a limited number of generating plants to many small-scale consumers. Renewable energy (particularly rooftop solar) requires a much more complex bi-directional grid where individuals are sometimes consumers and sometimes producers. ML regression algorithms can quickly analyse data, and automatically activate and deactivate components as the energy flows change.
Adverse Weather Conditions
Hailstorms, sandstorms, snow, and high winds can damage solar plants. Sandstorms and snow can considerably reduce energy output, but with a suitable ML algorithm (what the plant should be producing at this time if clean) robotic cleaners can be used economically. In terms of robotic cleaners, ML will be used to optimally utilize their battery life too. Hailstorms and high winds can damage solar panels permanently, ML impact prediction algorithms can be merged with panel positioning systems to minimize damage (e.g., vertical positioning of a panel in a hailstorm prevents impact damage).
In a more climate-friendly future, effectively managing grid stability comprising multiple distributed generating plants, and both small-scale local and large-scale national storage, will require sophisticated algorithms providing accurate prediction, fast adaption, and quick responses. Fortunately, all these requirements are met by ML algorithms. Development in this area is just beginning.
Author: Dr. Hugh Hind, CTO at Quadrical Ai
0 Comments on "Powering Solar Asset Management with Machine Learning"