Turning Big Data into Clean Electrons at NextEra
As prices drop and competition intensifies in the world of solar and wind, can generators use machine learning to maintain healthy margins?
The Renewable Paradox
Renewable energy is a victim of its own success. Solar and wind project costs have dropped precipitously in recent years and now regularly outcompete fossil fuels as the least expensive form of energy generation.[1] However, with this decrease in cost, the industry has become much more competitive and increasingly commoditized, leading to tight margins and high-profile bankruptcies, particularly in the solar project market.[2] Furthermore, renewables produce electricity intermittently, making a utility’s job of perfectly matching electricity supply and demand even more difficult. As more renewables are connected to the grid, the intermittency issue grows, and integration becomes increasingly challenging.[3] To be successful in this competitive space, renewable energy generators such as NextEra Energy are using machine learning to drive down costs and improve predictability of supply.
Optimizing Operations
NextEra claims to produce more electricity from solar and wind resources than any other company in the world.[4] This scale provides unprecedented access to operational data, relating the conditions of a solar or wind farm to its energy output. By processing this data in meaningful ways using techniques such as machine learning, the company can reduce costs and increase output, creating sustainable competitive advantage in a challenging field. NextEra is already working to apply two distinct applications of machine learning: optimizing operating parameters and performing predictive maintenance.
Setting operating parameters for its wind turbine fleet was once a manual process. Now, NextEra uses machine learning to dynamically adjust those parameters based on current conditions and continuously seek to maximize output.[5] The results have been outstanding: “We operate at $3 to $4 [per MWh] better, including availability and operating costs, on the wind side than anyone else in the country.”[6] With a fleet worth billions and slim margins, even small improvements in performance can create significant returns.
Predictive maintenance allows the operator to more accurately assess when an asset needs to be serviced, reducing maintenance costs and equipment outages. Here, machine learning supports the engineering team, according to the general manager of NextEra’s Advanced Data Systems team: “in the past [we] solely relied upon engineering knowledge to detect problems.…Machine learning recognizes patterns and anomalies within the data that may not be as easy for engineers to see.”[7] As a result, the company can perform maintenance based on the condition of the equipment, rather than at specified time intervals, meaning they can “stretch out the maintenance until the equipment needs it.”[8]
In the longer term, NextEra can use data to select the best sites for wind and solar projects, as with its purchase of wind forecasting company WindLogics.[9] Supporting that capability is a major area of machine learning research into forecasting solar and wind output using weather and other environmental data. A recent paper uses one such technique to predict wind turbine output with error of 1-5% over timespans ranging from an hour to a year in advance.[10] With improved forecasting, NextEra will be able to accurately bid on new projects and optimize performance of its existing assets.
A Clean Energy Future?
Looking forward, machine learning presents many more opportunities to generate further competitive advantage for the company. For example, forecasting algorithms can help Florida Power & Light, a subsidiary, better manage electricity generation. A more accurate forecast of renewable power generation would allow the utility to minimize its generation reserve, thereby reducing costs. On the project development side, a startup called PowerScout claims to have developed software that predicts consumer uptake of solar energy, making it easier for companies to market to the best prospects.[11] NextEra could use a similar concept to predict demand for its new solar or wind projects, perhaps by anticipating where utilities will need additional generation or storage and accounting for characteristics of potential sites.
However, the role of machine learning in creating sustainable competitive advantage for NextEra is unclear. As more operators realize the value of big data in managing their operations, will NextEra be able to sustain its advantageous cost position, or will others be able to reach the same level? More generally, will NextEra’s data advantage be enough to maintain its historically strong margins in a time of intense competitive pressure? (713 words)
References
[1] Silvio Marcacci, “Cheap Renewables Keep Pushing Fossil Fuels Further Away From Profitability – Despite Trump’s Efforts,” Forbes, January 23, 2018, https://www.forbes.com/sites/energyinnovation/2018/01/23/cheap-renewables-keep-pushing-fossil-fuels-further-away-from-profitability-despite-trumps-efforts, accessed November 2018.
[2] David Frankel, Aaron Perrine, and Dickon Pinner, “How solar energy can (finally) create value,” McKinsey Insights, October 2016, https://www.mckinsey.com/business-functions/sustainability-and-resource-productivity/our-insights/how-solar-energy-can-finally-create-value, accessed November 2018.
[3] Kasun S. Perera, Zeyar Aung, and Wei Lee Woon, “Machine Learning Techniques for Supporting Renewable Energy Generation and Integration: A Survey,” International Workshop on Data Analytics for Renewable Energy Integration: 96, Springer, accessed November 2018.
[4] NextEra Energy, Inc., “Renewable Energy,” http://www.nexteraenergy.com/sustainability/environment/renewable-energy.html, accessed November 2018.
[5] Stevan Jovanovic, “Innovation & Digitalization in Energy,” speech given at HBS Energy & Environment Symposium, Harvard Business School, Boston, MA, November 3, 2018.
[6] Jim Robo, CEO, remarks made at 2018 Wolfe Research Utilities & Energy Conference, New York, NY, October 2-3, 2018. From transcript provided by SeekingAlpha, https://seekingalpha.com/article/4209709-nextera-energy-inc-nee-ceo-jim-robo-2018-wolfe-research-utilities-and-energy-conference?page=4
[7] “Predicting the future of generation,” Energy Now, January 31, 2017, http://www.nexteraenergy.com/energynow/2017/0117/0117_PredictiveAnalytics.shtml, accessed November 2018.
[8] Ibid
[9] Russell Gold, “How a Florida Utility Became the Global King of Green Power,” The Wall Street Journal, June 18, 2018, https://www.wsj.com/articles/how-a-florida-utility-became-the-global-king-of-green-power-1529331001, accessed November 2018.
[10] Gul Muhammad Khan, Jawad Ali, and Sahibzada Ali Mahmud, “Wind power forecasting—An application of machine learning in renewable energy,” 2014 International Joint Conference on Neural Networks (IJCNN): 1137, IEEE, accessed November 2018.
[11] “PowerScout Brings Machine Learning to Consumer Energy Market,” PowerScout press release (Oakland, CA, September 14, 2016).
Featured image: Stephanie Sawyer, Getty Images, https://fortunedotcom.files.wordpress.com/2017/12/gettyimages-518650463.jpg, accessed November 2018.
In aviation, we often used “time before overhaul” (time based) standards for parts replacement. Conditions based standards often required labor intensive visual inspections or expensive automated vibrations analysis systems. I’m curious to know what upfront costs are involved to track this kind of data across equipment as large as a wind farm. Is this technology already in place to facilitate the machine learning?
While I see the role of machine learning in increasing the accuracy of predicting where the demand is, historically energy generation has been an industry that is extremely vulnerable to regulatory concerns which also play a major role in the construction and distribution of energy related assets such as where a plant is built. I am curious whether machine learning will simply provide an additional point of intelligent input into what are ultimately human decisions or whether there is a way to incorporate regulation as an idiosyncratic variable into the predictive model.
It’s great to see cost reduction and operational improvement in solar and wind beyond direct asset innovation. While improved material science enabling larger wind turbines is great, the extra improvement machine learning and other technologies offer provide compounding benefits which accelerate the trend toward renewables you discuss in the beginning of your article. It will be interesting to follow NextEra’s use of machine learning to see where else they apply it in the coming years as they strive to stay ahead of the competition.
Great read! It seems like NextEra has been able to leverage its volume to get a sufficient amount of data as input for their models. I wonder if this will deter smaller players to stay in this industry given their lack of data volume. Will big data as a competitive advantage only work for the biggest players in the industry, thus creating an oligopoly?
Awesome paper! I am curious where the data for the predictive maintenance comes from. As new windmills are developed with different technology, how long does it take for there to be enough data for the algorithm to provide value? Furthermore, since NextEra is collecting large amounts of windmill performance data (that few others are collecting), could a special relationship be developed with a windmill manufacturer in which the data is shared with an engineering team to optimize new windmill design? If so, perhaps this relationship could be contractually protected thus providing NextEra with another edge!
Great article. As someone who spent the pasts 5+ years in the renewable energy industry, I can attest that we are not always the most technologically advanced sector, so it is awesome to see that industry leaders like NextEra are finding ways to push the industry forward with machine learning. That said, as you point out, I think the biggest challenge facing renewable energy is its inherent intermittency and I am skeptical about machine learning’s ability to address that issue. Fundamentally, solving the intermittency challenge will require dramatic cost reductions in the upfront cost of energy storage and/or the construction of new transmission lines and I don’t see machine learning playing a role in either. (I would love to discuss if someone sees that differently!) But I do think machine learning can play an important role in optimizing a battery’s lifetime value by learning how to select the right charging/discharging opportunities each day. This is an extremely complicated optimization given that, among other things, electricity prices fluctuate in real time and are often unpredictable (and choosing to charge/discharge at one time of day means you forego the opportunity to charge/discharge at later time), so you can only truly know the best time to charge/discharge a battery in hindsight.
Thanks for writing this, Trey! I learned a lot.
My question after reading this is whether machine learning can do more than forecast when maintenance is needed. Can it, for example, toggle between renewable and non-renewable sources when the former are much more abundant because of favorable conditions (better wind conditions, higher levels of sunlight)? Perhaps my question is quite naive and electricity grids don’t actually work that way, but I am wondering if AI will take us into a world where it does more than flag things for maintenance but rather take action before humans can.