Xcel Energy: New challenges and ML solutions

Electric utilities are facing new challenges as they try to integrate solar and wind generation. Machine learning technologies may have some of the answers.

Renewable energy generation in the United States has grown nearly 12% per year in the last decade, and now accounts for ~10% of total generation[1]. To supply energy that is reliable, cheap, and low-carbon, electric utilities have to overcome a number of new challenges related to integrating renewable and traditional energy sources. For example, energy supplied from renewable sources varies with the weather and time of day, which requires the utility to be ready to supply power from storage or traditional sources on short notice and can lead to price volatility[2]. Balancing supply and demand will get even harder for the utility when some consumers own their own generation and storage assets – for example, in the form of rooftop solar panels and the battery in their electric car. Getting this balance wrong – for example, by under-forecasting the amount of solar energy production – is undesirable because it requires the utility to unnecessarily bring additional generation assets online on short notice, which in turn leads to higher prices and incremental pollution.

Going forward, utilities have several levers that they can pull to resolve these challenges: (1) improve the yield from generation assets; (2) improve electricity demand forecasting to enable better supply planning; (3) provide consumers with more options, both for reducing their use at high-demand times of day and for supplying excess energy from distributed generation and storage to other consumers; and (4) predict maintenance needs to reduce reliability issues[3]. Technologies that use machine learning are well suited to these challenges because the relationship between the information known and the variables the utility wants to control is complex and the training data is abundant and readily available.

The remainder of this essay explores some of the opportunities to apply machine learning to the problems that utilities are facing. We’ll consider the opportunities through the lens of Xcel Energy, which serves electricity and natural gas customers mainly in the Midwest. Xcel is piloting several technologies based on machine learning. First, to reduce emissions of nitrous oxide (a pollutant) at one of its plants, Xcel used data from a network of sensors in combination with a type of machine learning technology called an “artificial neural network.” The system can provide recommendations for how to operate the plant in real time to reduce NO pollution[4]. Second, since 2011, Xcel has collaborated with the National Oceanic and Atmospheric Association to more accurately forecast power output from wind turbines. Xcel reports that this program has avoid more than 250,000 tons of C02 emissions per year[5],[6]. Third, Xcel has become the first utility in the US to systematically use drones in conjunction with a machine learning algorithm to automatically inspect its infrastructure for potential failures2,[7]. The goal of this program is to pre-empt avoidable failures.

These use cases are promising, and there are still more opportunities that Xcel could explore. For illustration, we will consider two examples.

First, to improve the yield from power generation assets, GE introduced its “digital wind farm” solution. It uses a machine learning technology to simulate the way turbines and their environment would interact with the weather in a specific area. The solution would enable GE to customize the equipment design of each turbine based on its exact location and continuously optimize the way the turbines run[8]. GE believes it could increase efficiency of wind turbines by twenty percent.

Second, to further improve electricity demand forecasting to enable better supply planning, Google DeepMind entered a partnership with National Grid, which operates transmission infrastructure in the UK. DeepMind’s solution had previously been used to reduce electricity consumption at a Google datacenter by forty percent. It remains to be seen if the tool can be extended to an entire grid[9].

Third, to help consumers reduce their demand for electricity, a company called Bidgely is using machine learning to extend the benefits of smart electricity meters to homes that do not have them. Smart meters provide a number of benefits including detailed data about electricity usage and cost at the level of individual appliances and activities[10]. Consumers could use this data to more deliberately choose how much energy they use and at what times of day. Bidgely’s technology, which they call “Universal Disaggregation,” uses data from homes that have smart meters to estimate energy usage statistics for homes that do not[11].

More opportunities to use machine learning will arise as utilities continue to add renewable energy generation capacity and storage. How will Xcel Energy integrate decentralized energy storage – for example, using the batteries in electric buses? Can personalized recommendations for consumers actually encourage conservation at such low prices? Utilities will undergo dramatic changes in the next decade.

[1] U.S. Energy Information Administration, “Electricity Data Browser,” https://www.eia.gov/electricity/data/browser/, accessed November 2018.

[2] Frost and Sullivan, “Impact of Artificial Intelligence (AI) on Energy and Utilities, 2018,” https://cds-frost-com.prd1.ezproxy-prod.hbs.edu/p/71319/#!/ppt/c?id=ME1F-01-00-00-00&hq=machine%20learning%20utilities, accessed November 2018.

[3] Gavin Mooney, “10 ways Utilities can use AI and Machine Learning – right now,” https://blogs.sap.com/2018/06/28/10-ways-utilities-can-use-ai-and-machine-learning-right-now/, accessed November 2018.

[4] Connor Riffle, “What artificial intelligence means for sustainability,” https://www.greenbiz.com/article/what-artificial-intelligence-means-sustainability, accessed November 2018.

[5] Tadas Jucikas, “Artificial Intelligence and the future of energy,” https://medium.com/wepower/artificial-intelligence-and-the-future-of-energy-105ac6053de4, accessed November 2018.

[6] NOAA Research News, “Weather data from nation’s largest wind farms could improve U.S. models, forecasts,” https://research.noaa.gov/article/ArtMID/587/ArticleID/1445/Weather-data-from-nation%E2%80%99s-largest-wind-farms-could-improve-US-models-forecasts, accessed November 2018.

[7] Xcel Energy, “Drones,” https://www.xcelenergy.com/energy_portfolio/innovation/drones, accessed November 2018.

[8] GE Reports, “Wind in the Cloud? How the Digital Wind Farm Will Make Wind Power 20 Percent More Efficient,” https://www.ge.com/reports/post/119300678660/wind-in-the-cloud-how-the-digital-wind-farm-will-2/, accessed November 2018.

[9] Peter Maloney, “UK’s National Grid eyes Google AI unit to balance power supply and demand,” https://www.utilitydive.com/news/uks-national-grid-eyes-google-ai-unit-to-balance-power-supply-and-demand/437981/, accessed November 2018.

[10] California Public Utilities Commission, “The Benefits of Smart Meters,” http://www.cpuc.ca.gov/General.aspx?id=4853, accessed November 2018.

[11] Bidgely website, “Reaching 100% of Homes,” http://www.bidgely.com/blog/reaching-100-of-homes/, accessed November 2018.


Walmart Fights Fire with Fire: Traditional Retail in the Age of Machine Learning


Boeing: “Adding” to the Manufacturing Process

Leave a comment