AES enlists AI to boost its sustainable energy business

American utility and power company AES launched a renewable energy program in mid-2022 that is not only reducing its carbon footprint but adding wealth to its coffer.

To achieve that, the Arlington, Va.-based company, which claims to be the top-ranked supplier of renewable energy sales to corporations, turned to machine learning to help forecast renewable asset output, while establishing an automation framework for streamlining the company’s operations in servicing the renewable energy market.

The project, dubbed Farseer AI Generation Forecasting and Market Automation Program, was developed by a handful of AES data scientists in partnership with Google. The private cloud platform, which earned AES a 2024 CIO Award for IT leadership and innovation, combines the Farseer predictive analytics and machine learning tool for forecasting energy generation with AMART, a framework that works in tandem with Farseer to provide operational support by automating the complex and time-sensitive process of responding to fluctuating market demands.

AES is in its fifth year of partnership with Google — a unique relationship that sees Google as both a technology provider for AES and a customer.

“We build solutions for Google on every new workspace, and we consume their services,” says Alejandro Reyes, chief digital officer for AES Clean Energy, where he has worked for 14 years. As part of that partnership, Google operates AES’ private cloud platform in its data centers.

The Farseer and AMART program, which kicked off in July 2022, generated $3.4 million in its first year, contributed a $5.5 million bump in 2023, and the company predicts the analytics and machine learning platform’s contribution will increase to $8 million in 2024. Not bad for a sustainability initiative designed to reduce the $12.7B company’s carbon footprint to reach Paris Accord goals by 2050.

Fine-tuning forecasting with AI

Farseer and AMART are part of a larger concerted effort by AES to expand its investments in renewable energy. For example, the company bought Valcour Wind Energy’s six wind farms in New York and also builds and operates solar farms and storage systems in California, Arizona, and several other US states, as well as in Brazil and Argentina. Those investments come just as the company claims, in its 2023 annual report, that demand from corporate data centers in the US is expected to roughly double within the next three years as generative AI deployments expand.

But with the addition of more renewable energy to its portfolio, weather uncertainty becomes a greater challenge for AES. The Farseer machine learning model represents a major advancement for the company because it analyzes large amounts of historical weather data to predict wind farm output with far greater accuracy than in the past, Reyes says.

“As we iterated with Farseer, we’ve been able to make the model more accurate because the model is looking at the history and updates every day,” Reyes points out. “It proposes next-day generation and then we get the actuals, and the actuals go into the history and continues to further refine what the model is giving you the next day.”

AES’ use of machine learning demonstrates the significant value of AI in renewable energy forecasting, says John Villali, senior research director of energy insights at IDC.

“The increase in intermittent and unpredictable output for Distributed Energy Resources [DER] and renewable resources has increased the need for advanced AI-powered energy forecasting models,” Villali says. “AI-driven energy forecasting can help the power and utility sector in areas such as operations, trading, and integrated resource planning … and also support core functions such as long-term capital planning, short- and long-term load forecasts, economic dispatch of generating units, bidding strategies for capacity markets, demand response events, load shedding, load shifting, and more.”

(DER is a small-scale unit of power generation that operates locally and is connected to a larger power grid at the distribution level, Villali explains.)

Reyes notes that AES’ Farseer model can generate a forecast for any AES asset within the continental US. “Not only can we forecast for any asset, but we have numerous models — long term, short term, and intraday — which are picking up all the intricacies in weather that are specific to the location of the facility,” he says.

“We also generated each of these models from scratch, so they are not your standard black-box weather models used by most third-party forecasting companies,” he adds.

Sustainable precision

Once Farseer has provided an informed estimate of the megawatts of renewable energy AES will be able to offer the next day, the second aspect of the project, AMART, automates the process of fluctuating market demand to meet distribution demands in the most cost-effective manner, according to AES. Prior to implementing the AMART automation framework, the process was manual and heavily prone to human errors.

Finally, AES takes the data output of megawatts available and makes offers to the commercial market. “The cutoff time for the submission is 4 p.m., so all of the data is [generated] and automated in the morning, and we run the processes,” Reyes notes. “Before 4 p.m., we get the data to the commercial operations team.”

Reyes acknowledges the high value of the machine learning model as foundational in predicting and forecasting renewable energy on a daily basis.

“The model is running on our enterprise AES Google Cloud Platform environment, gathering actual generation data from our assets and weather data from different sources as an input, to then use in the machine learning model to generate the forecast as an output, which is then presented on a PowerBI dashboard to the traders and in other formats to the dispatchers so they can do their energy submittals,” Reyes tells CIO.com.

AES is expanding in the US and in Latin America and expects full deployment by the end of this year. Reyes says it may take some time for those customers to appreciate how well the machine learning model evolves.

“We are using the same model now simply trying to get data and trying to get people in those countries comfortable with the results because it is a journey,” Reyes says. “It’s a machine learning model but it has some learning. The model is adjusting the results based on what is proposed and what the actual reality [of renewable energy availability is], and then use that to give the next day’s forecast.”

According to Ember, a sustainability think tank based in England, renewable energy — particularly wind and solar — generated roughly 30% of global electricity in 2023, with solar accounting for 23% of the growth and 10% derived from wind generation. The US Energy Information Administration claims nonfossil fuel energy sources accounted for 21% of US energy consumption in 2022.

IDC’s Villali says the energy industry accounts for approximately 35% of total CO2 emissions globally — and there is more work to do.

“The energy transition is a global movement away from traditional fossil fuels such as coal, oil, and natural gas and toward cleaner forms of energy such as wind, solar, and energy storage,” Villali says. “The energy sector as a whole on average is targeting to become net-zero emissions in the 2031-2035 time frame.  This is an aggressive target that will need substantial global investment in renewable energy and much less reliance on fossil fuels to meet those goals.”

For AES, Farseer and AMART are already paying dividends.

“With more precision we see less deviations and are ultimately able to maximize our bottom line,” Reyes says, adding that much of that is due to Farseer’s ability to generate precise forecasts for AES assets when other models couldn’t generate forecasts at all.

“By improving renewable generation forecasting we are becoming a more efficient renewable energy company overall,” he says. “We are also utilizing AES’ expert knowledge in the renewable space instead of relying on third-party vendors,” which helps make the IT efforts more sustainable as well.