Anticipating energy demand using neural networks
Quick summary: In optimizing their operations to collectively make the best demand forecasting decisions using data, utilities are turning to neural networks.
When it comes to predicting the demand for electricity in the United States, a good starting strategy is to “follow the money”: The U.S. federal government has committed over $300 billion to supporting the nation’s energy sector. Energy macro-economics are being influenced by new and revised government policies:
• The U.S. government has set a goal of reaching 100 percent carbon pollution-free electricity by 2035, supported by initiatives such as a transition of the federal infrastructure to zero-emission vehicles and buildings, incentives for efficiency upgrades, electrification in private-sector buildings, and an electric vehicle (EV) charging infrastructure, just to name a few.
• The 2022 Infrastructure Investment and Jobs Act provided $3 billion in new funding to the existing Smart Grid Investment Grant Program to support the addition of smart grid functionality to distribution and transmission systems.
• On the state level, incentives are driving residential and commercial installations of wind, solar, and energy storage systems, as well as purchases of EVs.
The profile of U.S. energy consumption is changing rapidly, and utilities are responding to these macro events, particularly around demand forecasting, using advanced analytics. How do you optimize operations—workforce, physical assets, investments, and energy trading—to collectively make the best forecasting decisions using data? The answer lies in neural networks.
Neural networks: the future of demand forecasting
A neural network is a series of algorithms that analyze structured and unstructured data from hundreds of sources to find patterns and identify trends. These solutions have been successfully leveraged in the financial services industry, in areas such as high-frequency trading, and are gaining traction among utility companies looking to fine-tune their demand forecasts.
Neural networks can support short- and long-term electricity demand forecasting by analyzing large amounts of data, both structured and unstructured, from a broad variety of sources (more on this below).
Much of the work on neural networks to date has been in the area of supervised learning, in which a system is “told” what to look for. Advanced neural networks are created when the model recognizes patterns on its own and starts to modify for those patterns—also known as unsupervised learning. Networks capable of unsupervised learning offer the advantage of being able to generate energy demand predictions that no one may have anticipated—another potential win for utilities.
The exciting paradigm of demand forecasting
To plan the right amount of capacity for any given time span, utilities must invest in forecasting demand accurately. If forecasts are too low, the imbalance could result in brownouts or blackouts—like the ones amid the February 2021 winter storm, which caused over 5 million inhabitants across the United States to lose power. If forecasts overshoot the mark, utilities (and, inevitably, you and I) wind up paying for excess capacity (regardless of whether we remember to tell Alexa to switch off the garage lights).
Major blackouts—those lasting at least 1 hour and impacting 50,000 or more customers—in the United States increased by more than 60% since 2015.
Electricity demand forecasts have traditionally relied heavily on historical data, including demographics, weather patterns, and consumption records. However, the following three trends are making these methods obsolete (or at least insufficient) and requiring a more sophisticated approach.
Trend #1: Growth in on-site renewable energy sources
Driven by concerns over climate change and encouraged by falling prices as well as government incentive programs, more residential and commercial power consumers are opting for on-site renewable electricity generation.
Solar: In 2021, the residential solar market recorded its fifth consecutive record year, growing 30 percent over 2020.
Wind: The U.S. distributed wind sector added 14.7 megawatts of new distributed wind energy capacity in 2021, with 1,493 new wind turbines installed across 11 states.
Energy Storage: Q4 of 2021 was the strongest quarter to date for residential energy storage, with 123 megawatts installed.
While this development is good news from an environmental standpoint, it also complicates utilities’ efforts to forecast demand accurately. Consumers with on-site wind and solar generation rely on the grid only as a supplement to their own sources. As generation capacity varies from system to system—not to mention variations due to weather fluctuations—utilities can run into difficulties trying to forecast the demand from these customers using traditional approaches.
Neural networks can help utilities incorporate on-site renewables data into their demand forecasting by analyzing
• Solar, wind, and energy storage market reports
• Details of federal and state incentive programs
• Consumer sentiments regarding grid reliability
Trend #2: Rise of the smart grid
While the idea of a smart grid is hardly new—the term has been around since 2007—the latest advances in smart-grid technologies are offering utilities a valuable source of data in their mission to forecast demand accurately.
Smart grid technologies provide utility providers with a continuous data feed that captures demand patterns at the individual household level. For consumers with on-site wind or solar generation, utilities can capture in real time how much power they are generating, consuming, buying from the grid, and selling back to it.
Thanks to the evolving smart grid, utilities have at their disposal a continuous feed of data capable of fostering increasingly accurate demand forecasts. Now the challenge is wrangling this tsunami of data into actionable insights that can inform data-driven decision making—a task ideally suited to neural networks.
Trend #3: Increasing electrification
Spurred by an increase in the number of electricity-consuming devices and a shift away from earlier power sources such as natural gas, electrical power consumption is in growth mode:
• Registrations of EVs skyrocketed in the first quarter of 2022, even as the overall auto market declined, and many businesses—including Amazon, Walmart, and FedEx—are electrifying their vehicle fleets.
• Thousands of cryptocurrency organizations based in the United States are consuming huge amounts of energy. Bitcoin, to cite just one example, uses about 150 terawatt-hours of electricity each year—more than the entire country of Argentina.
• Many consumers are swapping out fossil fuel–powered household devices such as lawn mowers and home heating systems for electricity-powered alternatives.
Growing electrification is changing the shape of electricity demand in the United States, creating a challenge for utilities looking to develop accurate predictions and diminishing the effectiveness of traditional forecasting methods.
Using neural networks, utilities can fine-tune their demand forecasting by incorporating an array of data sources surrounding electrification, including
• Trends in manufacturing, retail, and other energy-intensive industries
• Sales and registrations of electric vehicles
• Real estate development trends
New trends, new approaches
Our world is growing more complex and continues to rapidly change. Energy supply and demand is no different. While historical data still has its place in short- and long-term demand forecasting, it is no longer sufficient on its own. Fortunately, intelligent technologies such as smart meters and sensors provide the data utilities need for more targeted forecasting, and neural networks deliver the analytical capabilities for spinning this data into strategic decision-driving insights. Combining these powerful innovations enables utilities to keep up with the pace of change, enhancing their ability to precisely deliver the amount of power consumers demand—when and where they need it.
Some utility providers may still be reluctant to implement advanced digital solutions in their demand forecasting. However, due to the complex and evolving profile of electricity demand, the decision to stick with the status quo is a risky one. If utilities are to avoid the costs of over- and under-forecasting demand, intelligent data sources and neural networks are the way to go. The sooner they get on board with these solutions, the sooner they can start reaping the benefits.
Amit Unadkat is a Senior Manager of Digital Transformation with extensive experience in robotic process automation, virtual assistants, business process optimization, and technical product management. In 2021 he received Built In’s Tech Innovator Award for his work in automation and was recognized as a Rising Star by Consulting Magazine.